A non-intrusive load identification method and device based on weight-based model body differential field

Abstract

The application provides a non-intrusive load identification method and device based on weight model difference field, and relates to the technical field of intelligent power grids and non-intrusive load identification. The method comprises the following steps: performing event detection on a data set to obtain a current sequence, decomposing and extracting a reactive current sequence, and constructing a weight of the reactive current sequence according to a harmonic coefficient sequence; obtaining a hybrid subsequence with rich time-frequency characteristics by performing variational mode decomposition and wavelet packet decomposition on the current sequence; constructing a two-dimensional RGB three-channel load feature map for the current sequence, the hybrid subsequence and the weighted current sequence by model difference; and building a load identification model based on a ResSENet architecture to perform non-intrusive load identification. The application introduces a new imaging method, converts the current sequence, the reactive current sequence and the time-frequency hybrid sequence into the RGB three channels of a two-dimensional picture, and uses the sum of the harmonic coefficients as the weight of the reactive current channel to improve the feature correlation between the channels.

Description

Technical Field

[0001] This invention relates to the field of smart grid and non-intrusive load identification technology, and in particular to a non-intrusive load identification method and device based on weighted modulus differential field. Background Technology

[0002] NILM (Non-Intrusive Load Monitoring) is a technology that analyzes total electricity meter data for a specific area to obtain information about the electrical loads within that area. This information includes the number, type, operating status, and energy consumption of the loads. Compared to traditional intrusive monitoring methods, NILM does not require the installation of a separate monitoring device for each electrical device, thus saving costs and reducing installation and maintenance hassles. As a low-cost, easy-to-deploy, and privacy-protecting monitoring method, NILM has broad application prospects.

[0003] NILM research primarily focuses on algorithm development and dataset construction. Early research was mainly based on traditional methods, such as Hart's method of identifying appliance types using low-speed data based on power variations. With the development of deep learning, more and more research is beginning to use deep learning algorithms to improve recognition accuracy. For example, methods based on Faster R-CNN perform exceptionally well on small sample data, significantly improving recognition accuracy and recall.

[0004] Non-intrusive load detection mainly includes two tasks: load decomposition and load identification. Solving the load identification problem primarily involves three steps: main meter data acquisition and preprocessing, detection of appliance tripping events, and extraction of load features and construction of an identification model to distinguish load categories. The acquired data types generally fall into two categories: high-frequency datasets and low-frequency datasets. Low-frequency acquisition instruments often collect power signals; however, due to the limited features contained in low-frequency data, correctly identifying the operating category of appliances is difficult. Therefore, using high-frequency datasets for load identification is currently the mainstream approach. Existing research generally extracts features such as reactive power, active power, harmonic information, and current and voltage waveforms for load identification. Typically, features that can distinguish different appliances and their states and can be used for load identification are called load fingerprints, also known as load signatures. With the rapid development of deep learning, research using image representations as load signatures has made significant progress.

[0005] Currently, in the field of non-invasive payload recognition, the main methods for constructing payload fingerprints using image representation include: VI (Voltage-Current) trajectories, Gram angle fields, Markov transformation fields, and recursive graphs. Based on research and understanding, the principles and main technologies disclosed for these four image representation methods are as follows:

[0006] VI tracing, a technique for mapping voltage and current data to two-dimensional images, was first introduced in 2007 by the paper (Lam H, Fung G, Lee W. A Novel Method to Construct Taxonomy Electrical Appliances Based on Load Signatures of [J / OL]. IEEE Transactions on Consumer Electronics, 2007, 53(2): 653-660.). Since then, numerous studies have focused on this technique. Among them, Xie Yang et al. (A Non-Intrusive Load Identification Method Based on VI Tracing Color Encoding [J]. Automation of Electric Power Systems, 2021.) first used k-means to perform preliminary classification based on the reactive and active power characteristics of the load. For appliances that failed to be classified, color-coded VI tracings were constructed, and finally, AlexNet was used to achieve fine identification. The paper (De Baets L, Dhaene T, Deschrijver D, et al. VI-Based Appliance Classification Using Aggregated Power Consumption Data[C / OL] / / 2018 IEEE International Conference on SmartComputing (SMARTCOMP). Taormina: IEEE, 2018: 179-186.) introduces two classification techniques: one is to use a convolutional neural network to classify VI trajectories, and the other is to use a random forest classifier built with the elliptic Fourier descriptor of the VI trajectory. The paper (Liu Y, Wang X, You W. Non-Intrusive Load Monitoring by Voltage–Current Trajectory Enabled Transfer Learning[J / OL]. IEEE Transactions on Smart Grid, 2019, 10(5): 5609-5619.) innovatively applies color coding to map the trajectory direction, power features, and VI trajectory pixel mean to the HSV color space, thereby enhancing the recognizability of VI trajectories, and combines transfer learning to use a pre-trained AlexNet for classification.To address the issues of missing information and insufficient pixel utilization in VI trajectories, Wang et al. (WANG S, CHEN H, GUO L, et al. Non-intrusive load identification based on the improved voltage-current trajectory with discrete color encoding background and deep-forest classifier[J / OL]. Energy and Buildings, 2021, 244: 111043.) introduced motion and momentum information through color encoding and proposed a data balancing algorithm based on PixelCNN++. Compared with traditional convolutional classifiers, it achieved faster and more accurate recognition within the deep forest classification framework. However, due to the greater intra-class diversity brought about by multi-state electrical appliances, current research based on VI trajectories generally struggles to effectively identify multi-state electrical appliances, especially as the number of categories of multi-state electrical appliances increases, resulting in a significant decline in recognition performance. This has led to a scarcity of innovative research based on VI trajectories.

[0007] Gramian Angular Field (GAF) has the ability to capture temporal correlations between points within a sequence. In the paper (Chen J, Wang X, Zhang X, et al. Temporal and Spectral Feature Learning With Two-Stream Convolutional Neural Networks for Appliance Recognition in NILM[J / OL]. IEEE Transactions on Smart Grid, 2022, 13(1): 762-772.), the authors used GAF ​​to construct temporal and spectral load feature maps and designed a dual-channel convolutional network framework based on this, which extracts features from these two types of load feature maps respectively. At the same time, in order to alleviate the difficulty in effectively distinguishing load categories, especially for multi-state appliances, affinity propagation clustering algorithm is used in the framework to improve classification performance. However, in current research on load identification using GAF, the effective differential features extracted solely from time-series current sequences are insufficient. In addition, the normalization process in GAF inevitably involves a certain degree of information loss.

[0008] MTF (Markov Transition Field) maps time-series data into a matrix whose elements depict the predefined transition probabilities between different quantiles. Qu et al. (A residualconvolutional neural network with multi-block for appliance recognition innon-intrusive load identification[J / OL]. Energy and Buildings, 2023, 281:112749.) first generated three two-dimensional matrices: a Gram angle field converted from current, an MTF constructed based on data points in the VI trajectory, and a weighted VI trajectory diagram. These three two-dimensional matrices were then fused to form the entire load signature. The pre-defined number of quantiles, denoted as R, in the MTF affects the extracted time-domain features, but currently, there is a lack of effective means to determine this number. Therefore, the application of MTF in the field of non-intrusive load identification is relatively limited.

[0009] Recurrence Plot (RG) generates a binary image representation by measuring the distance between instances in the input sequence, but this binarization process is accompanied by information loss. To address this issue, Faustine et al. (Faustine A, Pereira L. Improved Appliance Classification in Non-Intrusive Load Monitoring Using Weighted Recurrence Graph and Convolutional Neural Networks[J / OL].Energies, 2020, 13(13): 3374.) introduced Weighted Recurrence Plot (WRG), which mitigates information loss by applying a threshold to the output. However, constructing a recurrence graph requires a predefined threshold. Therefore, the paper (Faustine A, Pereira L, Klemenjak C. Adaptive Weighted Recurrence Graphs for Appliance Recognition in Non-Intrusive Load Monitoring[J / OL]. IEEE Transactions on Smart Grid, 2021, 12(1): 398-406.) further proposed an adaptive weighted recurrence graph method, treating the threshold as a learnable hyperparameter that is optimized along with the network. Although existing literature attempts to optimize this, the performance gains achieved are relatively limited due to the lack of an effective threshold determination method.

[0010] The aforementioned four image-based methods in the field of load identification either extract single features from a single sequence or have inherent limitations that make it difficult to effectively extract differentiated features. Furthermore, existing methods for constructing multi-channel two-dimensional images as load fingerprints mainly focus on fusing multiple image-based methods for constructing load signatures. These fusion methods are mostly based on intuitive analysis, resulting in weak feature correlations between channels in the fused image and introducing significant uncertainty. This is a major factor limiting their ability to correctly identify appliance categories. Summary of the Invention

[0011] To address the limitations of existing image-based methods in load identification, which either extract single features from a single sequence, thus hindering the effective extraction of differentiated features, or rely on intuitive analysis to directly mix multiple different construction methods or load features, resulting in weak feature correlations between channels of the fused image and significant uncertainty, this invention provides a non-invasive load identification method and apparatus based on weighted phantom difference fields. The technical solution is as follows:

[0012] On the one hand, a non-invasive load identification method based on weighted phantom differential fields is provided. This method is implemented by a non-invasive load identification device and includes:

[0013] S1. Obtain the high-frequency public dataset of electricity meter data, perform event detection on the high-frequency public dataset, obtain the current sequence that marks the event jump, decompose and extract the current sequence to obtain the reactive current sequence, and construct the weight of the reactive current sequence based on the harmonic coefficient sequence.

[0014] S2. By performing variational mode decomposition and wavelet packet decomposition on the current sequence, a mixed subsequence with rich time-frequency features is obtained. A two-dimensional RGB three-channel load feature map is constructed from the current sequence, the mixed subsequence and the weighted current sequence through the modal difference field.

[0015] S3. Build a load identification model based on the ResSENet architecture, input the two-dimensional load feature map into the load identification model, and obtain non-intrusive load identification results.

[0016] Optionally, event detection on high-frequency public datasets in S1 includes:

[0017] The change in active power is used as the criterion for detecting whether an event transition has occurred to perform event detection on a high-frequency public dataset.

[0018] Optionally, the active power can be calculated as shown in equation (1) below:

[0019] (1)

[0020] In the formula, Indicates active power. This represents the number of data points within a given period. This represents the steady-state current over one cycle. It represents the steady-state voltage over one cycle.

[0021] The current sequence is shown in equation (2) below:

[0022] (2)

[0023] In the formula, The current sequence that indicates the transition to the flag event. It represents the complete current cycle intercepted from negative to positive at the zero-crossing point of the voltage in a sequence that is in a steady state before the event. It represents the complete current cycle intercepted from negative to positive at the zero point of the voltage in the steady-state sequence after the event.

[0024] Optionally, the harmonic coefficient sequence in S1 is as shown in equation (3) below:

[0025] (3)

[0026] In the formula, Represents the harmonic coefficient sequence. , Represents integers, Indicates a time period. Indicates time period Periodic functions within, Represents the imaginary unit. In the real number field The frequency on, Indicates a time step.

[0027] Optionally, the weights in S1 for constructing the reactive current sequence based on the harmonic coefficient sequence include:

[0028] The harmonic coefficient sequence is obtained by fast Fourier transform, and the sum of the first to fifth order harmonic coefficient sequences is used as the weight of the reactive current sequence.

[0029] Optionally, in S2, a mixed subsequence rich in time-frequency features is obtained by performing variational mode decomposition and wavelet packet decomposition on the current sequence. A two-dimensional RGB three-channel load feature map is constructed from the current sequence, the mixed subsequence, and the weighted current sequence using a modal difference field, including:

[0030] Multiple coefficient sets and components are obtained by performing variational mode decomposition and wavelet packet decomposition on the current sequence. The coefficient sets and components are then pieced together to obtain a hybrid subsequence with rich time-frequency characteristics.

[0031] The current sequence, reactive current sequence, and mixed subsequence are reduced in dimensionality using a piecewise aggregation approximation algorithm and the feature space is unified using max-min normalization to obtain the normalized current sequence, reactive current sequence, and mixed subsequence.

[0032] The normalized current sequence, reactive current sequence, and mixed subsequence are transformed into a two-dimensional matrix using modal difference field, forming the RGB three channels of a two-dimensional image. The weights of the reactive current sequence are multiplied by all values ​​of the reactive current sequence channels to obtain a two-dimensional RGB three-channel load feature map.

[0033] Optionally, the load identification model in S3 includes: a 3×3 convolutional layer, a first normalization layer, a first convolutional module, a first downsampling module, a second convolutional module, a second downsampling module, a third convolutional module, a third downsampling module, a fourth convolutional module, a global average pooling layer, a second normalization layer, and a linear layer.

[0034] The convolution module includes a 3×3 convolutional layer, Gaussian error linear units, and batch normalization.

[0035] The downsampling module includes a first branch, a second branch, a 1×1 convolutional layer, and a 3×3 convolutional layer; the first branch includes an average pooling layer, a ReLU-based linear layer, a Sigmoid-based linear layer, and a 1×1 convolutional layer; the second branch includes a 3×3 depthwise separable convolutional layer and a 1×1 convolutional layer.

[0036] On the other hand, a non-invasive load identification device based on a weighted phantom differential field is provided. This device is applied to a non-invasive load identification method based on a weighted phantom differential field. The device includes:

[0037] The event detection and data preprocessing module is used to acquire high-frequency public datasets of electricity meter data, perform event detection on the high-frequency public datasets, obtain current sequences that mark event transitions, decompose and extract the current sequences to obtain reactive current sequences, and construct weights for the reactive current sequences based on the harmonic coefficient sequences.

[0038] The RGB three-channel fusion feature construction module is used to obtain a mixed subsequence with rich time-frequency features by performing variational mode decomposition and wavelet packet decomposition on the current sequence, and construct a two-dimensional RGB three-channel load feature map by using the modal difference field on the current sequence, the mixed subsequence and the weighted current sequence.

[0039] The load identification model building module is used to build a load identification model based on the ResSENet architecture. It inputs a two-dimensional load feature map into the load identification model to obtain non-intrusive load identification results.

[0040] Optionally, the event detection and data preprocessing module is further used for:

[0041] The change in active power is used as the criterion for detecting whether an event transition has occurred to perform event detection on a high-frequency public dataset.

[0042] Optionally, the active power can be calculated as shown in equation (1) below:

[0043] (1)

[0044] In the formula, Indicates active power. This represents the number of data points within a given period. This represents the steady-state current over one cycle. It represents the steady-state voltage over one cycle.

[0045] The current sequence is shown in equation (2) below:

[0046] (2)

[0047] In the formula, The current sequence that indicates the transition to the flag event. It represents the complete current cycle intercepted from negative to positive at the zero-crossing point of the voltage in a sequence that is in a steady state before the event. It represents the complete current cycle intercepted from negative to positive at the zero point of the voltage in the steady-state sequence after the event.

[0048] Alternatively, the harmonic coefficient sequence is as shown in equation (3) below:

[0049] (3)

[0050] In the formula, Represents the harmonic coefficient sequence. , Represents integers, Indicates a time period. Indicates time period Periodic functions within, Represents the imaginary unit. In the real number field The frequency on, Indicates a time step.

[0051] Optionally, the event detection and data preprocessing module is further used for:

[0052] The harmonic coefficient sequence is obtained by fast Fourier transform, and the sum of the first to fifth order harmonic coefficient sequences is used as the weight of the reactive current sequence.

[0053] Optionally, the RGB three-channel fusion feature construction module is further used for:

[0054] Multiple coefficient sets and components are obtained by performing variational mode decomposition and wavelet packet decomposition on the current sequence. The coefficient sets and components are then pieced together to obtain a hybrid subsequence with rich time-frequency characteristics.

[0055] The current sequence, reactive current sequence, and mixed subsequence are reduced in dimensionality using a piecewise aggregation approximation algorithm and the feature space is unified using max-min normalization to obtain the normalized current sequence, reactive current sequence, and mixed subsequence.

[0056] The normalized current sequence, reactive current sequence, and mixed subsequence are transformed into a two-dimensional matrix using modal difference field, forming the RGB three channels of a two-dimensional image. The weights of the reactive current sequence are multiplied by all values ​​of the reactive current sequence channels to obtain a two-dimensional RGB three-channel load feature map.

[0057] Optionally, the load identification model includes: a 3×3 convolutional layer, a first normalization layer, a first convolutional module, a first downsampling module, a second convolutional module, a second downsampling module, a third convolutional module, a third downsampling module, a fourth convolutional module, a global average pooling layer, a second normalization layer, and a linear layer.

[0058] The convolution module includes a 3×3 convolutional layer, Gaussian error linear units, and batch normalization.

[0059] The downsampling module includes a first branch, a second branch, a 1×1 convolutional layer, and a 3×3 convolutional layer; the first branch includes an average pooling layer, a ReLU-based linear layer, a Sigmoid-based linear layer, and a 1×1 convolutional layer; the second branch includes a 3×3 depthwise separable convolutional layer and a 1×1 convolutional layer.

[0060] On the other hand, a non-invasive load identification device is provided, the non-invasive load identification device comprising: a processor; a memory storing computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, any one of the above-described non-invasive load identification methods based on weighted phantom difference fields is implemented.

[0061] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described non-intrusive load identification methods based on weighted modulus difference fields.

[0062] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0063] This invention proposes a novel method for constructing load fingerprints. For the first time, a weighted modal differential field is proposed to encode the current, reactive current, and time-frequency characteristics of the current into three RGB channels as load fingerprints for load identification.

[0064] We propose a load identification architecture, ResSENet, based on the design concept of the swin transformer. Compared with existing load identification methods, the model has less training time and a higher recognition rate. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a flowchart of a non-intrusive load identification method based on a weighted modulus difference field provided in an embodiment of the present invention;

[0067] Figure 2 This is a current sequence diagram of one cycle of the steady state during the operation of an incandescent lamp, provided in an embodiment of the present invention.

[0068] Figure 3 This is a current sequence diagram of one cycle in the steady state of a laptop computer during operation, provided in an embodiment of the present invention.

[0069] Figure 4 This is a mixed subsequence diagram of incandescent lamps obtained by wavelet packet decomposition and variational mode decomposition provided in the embodiments of the present invention;

[0070] Figure 5 This is a hybrid subsequence diagram of a laptop computer obtained by wavelet packet decomposition and variational mode decomposition provided in the embodiments of the present invention;

[0071] Figure 6 This is a diagram illustrating the process of constructing a two-dimensional load characteristic of a weighted modulus differential field, provided in an embodiment of the present invention.

[0072] Figure 7 This is a diagram of the proposed fully convolutional ResSENet architecture provided in this embodiment of the invention;

[0073] Figure 8 This is a schematic diagram of the SE attention mechanism provided in an embodiment of the present invention;

[0074] Figure 9 This is a block diagram of a non-intrusive load identification device based on a weighted modulus differential field provided in an embodiment of the present invention;

[0075] Figure 10 This is a schematic diagram of the structure of a non-invasive load identification device provided in an embodiment of the present invention. Detailed Implementation

[0076] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0077] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0078] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0079] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0080] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0081] This invention provides a non-invasive load identification method based on a weighted modulus difference field. This method can be implemented using a non-invasive load identification device, which can be a terminal or a server. Figure 1 The flowchart shown is for a non-intrusive load identification method based on weighted phantom differential fields. The processing flow of this method may include the following steps:

[0082] S1. Obtain the high-frequency public dataset of electricity meter data, perform event detection on the high-frequency public dataset, obtain the current sequence that marks the event jump, decompose and extract the current sequence to obtain the reactive current sequence, and construct the weight of the reactive current sequence based on the harmonic coefficient sequence.

[0083] In one feasible implementation, the PLAID and WHITED datasets are high-frequency publicly available datasets containing the transitions of electrical appliances from off to on and from one state to another. These transitions are typically referred to as event transitions, and event transitions are used to characterize the changing operating modes of electrical appliances. Therefore, event detection is required on the datasets. Currently, many studies use different evaluation criteria to determine the occurrence of event transitions, such as power changes and RMS current values. In high-frequency sampled data, it is generally assumed that only one state changes at any given time. This invention uses the change in active power as the criterion for detecting whether an event transition has occurred, where power can be calculated using the following formula:

[0084] (1)

[0085] In the formula, Indicates active power. This represents the number of data points within a given period. This represents the steady-state current over one cycle. It represents the steady-state voltage over one cycle.

[0086] Furthermore, after events in the time series data are effectively detected, a complete current cycle is extracted from the zero-crossing point of the voltage in the steady-state sequence before and after the event, from negative to positive, and denoted as follows: , Therefore, the current that can indicate a state transition is denoted as:

[0087] (2)

[0088] Furthermore, the symmetry, color coding, and background mapping of VI trajectories: For multi-state appliances (also known as Type II appliances), they have multiple operating states, such as washing, spin-drying, and standby modes of washing machines. These multi-state appliances bring greater intra-class diversity (IACV). At the same time, there are similar working modes among multi-state appliances, such as the heating mode of an air conditioner and the heating and blowing mode of a hair dryer. This will make it difficult to distinguish similar load fingerprints. Therefore, how to effectively distinguish multi-state appliances is a major challenge in load identification. According to Fryze's power theory (Teshome DF, Huang TD, Lian K L. Distinctive load feature extraction based on Fryze's time-domain power theory[J]. IEEE Power and Energy Technology Systems Journal, 2016, 3(2): 60-70.), the current signal can be decomposed into reactive and active components. The reactive component contains non-resistive information. Therefore, in this invention, the reactive component of the current is extracted as part of the load characteristics. The decomposition of the current signal in Fryze's power theory can be expressed by the following equation:

[0089] (3)

[0090] in and These represent active and reactive components, respectively. This is the effective value of the periodic voltage.

[0091] Fryze's power theory can represent reactive and active components as non-resistive and resistive information, which can further enhance the differentiated characteristics of WMDF.

[0092] If the image contains richer features, the neural network can make more informed judgments by learning these features. Therefore, this invention introduces harmonic features as weights for reactive current for the first time, enhancing the network's ability to uncover potential features and providing greater scope for feature learning. Loads are mainly divided into linear and nonlinear loads, and higher-order harmonic features provide linear information related to electrical appliances. Therefore, the introduction of harmonic features allows the VI trajectory to carry more information, including coefficients. The sequence can be represented as:

[0093] (4)

[0094] In the formula, Represents the harmonic coefficient sequence. , Represents integers, Indicates a time period. Indicates time period Periodic functions within, Represents the imaginary unit. In the real number field The frequency on, The time step is indicated. Harmonics are larger and more characteristic than even harmonics, and their contribution to characteristic differentiation is small above the fifth harmonic. Therefore, after the periodic current is processed by FFT (Fast Fourier Transform), this invention uses the sum of the first to fifth harmonic coefficients as the weight of the reactive current.

[0095] S2. By performing variational mode decomposition and wavelet packet decomposition on the current sequence, a mixed subsequence with rich time-frequency features is obtained. A two-dimensional RGB three-channel load feature map is constructed from the current sequence, the mixed subsequence and the weighted current sequence through the modal difference field.

[0096] Optionally, step S2 above may include the following steps S21-S23:

[0097] S21. Multiple coefficient sets and components are obtained by performing variational mode decomposition and wavelet packet decomposition on the current sequence. The coefficient sets and components are then pieced together to obtain a mixed subsequence with rich time-frequency characteristics.

[0098] S22. The current sequence, reactive current sequence, and mixed subsequence are reduced in dimensionality using a piecewise aggregation approximation algorithm and the feature space is unified using max-min normalization to obtain the normalized current sequence, reactive current sequence, and mixed subsequence.

[0099] S23. Use the modal difference field to convert the normalized current sequence, reactive current sequence and mixed subsequence into a two-dimensional matrix to form the RGB three channels of the two-dimensional image. Multiply the weight of the reactive current sequence with all the values ​​of the reactive current sequence channels to obtain the two-dimensional RGB three-channel load feature map.

[0100] In one feasible implementation, for the current sequence By using wavelet packet decomposition and variational mode decomposition, multiple sets of coefficients and components are obtained. These sets of coefficients and components are then pieced together into a sequence, which constitutes a mixed subsequence rich in frequency features. A diagram illustrating the acquisition of subsequences is shown below. Figures 2-5 Then, the current sequence () is analyzed using a modal differential field (MDF). ), reactive current sequence ( ), and mixed subsequences with rich frequency features These three time-series data were converted into two-dimensional matrices to form the RGB three channels of the image. Since these three sequences are all high-frequency data and have inconsistent lengths, PAA (Piecewise Aggregation Approximation) was used to reduce the dimensionality of the three sequences before constructing the two-dimensional image features. This not only reduces the time required to construct the two-dimensional image but also ensures that the three sequences produce features at the same scale. Then, min-max normalization was used to unify the feature space.

[0101] Furthermore, loads are categorized into linear and nonlinear loads. Linear loads can be resistive, capacitive, or inductive electrical appliances. The reactive component of the current provides non-resistive information, while higher-order harmonic characteristics provide information on the load's linearity. Therefore, harmonic characteristics are used as the reactive current sequence. The weight of the channel being converted is calculated as the sum of the harmonic coefficients multiplied by all values ​​in that channel. Therefore, this method of constructing load characteristics is called a weighted modal differential field. The overall framework for constructing a two-dimensional three-channel image load is as follows: Figure 6 .

[0102] Generate current sequence through event detection ( ), reactive current sequence ( ), mixed subsequences ( These three time series data are used to construct two-dimensional three-channel features, transforming the task of load identification into the task of image classification.

[0103] This invention introduces a novel image visualization method: Weighted Modulus Differential Field (WMDF). By converting current sequences, reactive current sequences, and time-frequency mixed sequences into three RGB channels of a two-dimensional image, and using the sum of harmonic coefficients as the weight for the reactive current channel, the feature correlation between channels is enhanced.

[0104] S3. Build a load identification model based on the ResSENet architecture, input the two-dimensional load feature map into the load identification model, and obtain non-intrusive load identification results.

[0105] In one feasible implementation, the present invention proposes a fully convolutional network architecture (ResSENet) with an embedded SE (Squeeze and Excitation) attention mechanism to perform load identification.

[0106] Specifically, in recent years, visual transformers have continuously challenged the dominance of convolutional neural networks (CNNs) in computer vision, especially the landmark Swing transformer model. This invention proposes a pure convolutional network architecture based on the Swing transformer design concept, exploring the greater potential of CNNs in non-intrusive payload recognition. The proposed network architecture, ResSENet, is as follows: Figure 7 In this context, Input represents the input layer, LN represents layer normalization, Conv represents a convolutional layer, GELU, ReLU, and Sigmoid represent different activation functions, BN represents batch normalization, Downsampling represents a downsampling module, Global Avg pool represents global average pooling, and Linear represents a linear layer. The three core design principles are as follows:

[0107] Stacking ratio of feature learning layers and output size of each stage: The Swin transformer provides a new feature learning mode by introducing an offset window mechanism and can serve as a general visual architecture in the field of computer vision. A detailed analysis of the internal structure of the Swin transformer reveals that the proposed ResSENet architecture aligns as closely as possible with the scaling ratio of the Swin transformer modules (2, 2, 6, 2) and the output size of each stage. Since the classification model built by the Swin transformer is mainly based on images of size 224*224 or 384*384, while the two-dimensional load feature map generated by this invention is 64*64, only three downsampling stages are set. Simultaneously, the stacking ratio of the convolutional modules is set to 1:1:3:1, making the model more suitable for the load recognition task of this invention.

[0108] GELU followed by Batch Normalization (BN): Research shows that activation functions are very effective in improving the learning performance and stability of networks. Currently, ReLU (Rectified Linear Unit) and GELU (Gaussian Error Linear Units) are commonly used nonlinear activation functions and can solve the gradient vanishing problem. ReLU typically uses a ramp function. ReLU is considered superior to the Sigmoid activation function in terms of computational efficiency and performance because it introduces greater sparsity into neural networks without involving costly exponential operations. However, ReLU cannot avoid the gradient explosion problem and also introduces the death ReLU problem. Unlike the ReLU family of activation functions (ELU, Leaky ReLU, SELU), GELU introduces the idea of ​​random regularization into the activation function, improving the robustness of the training process. It also uses the probability distribution of the input to determine whether to retain or deactivate the current neuron. Therefore, it is considered an activation function that combines the ideas of ReLU and Dropout.

[0109] During training, as the network deepens, the input distribution of the hidden layers gradually changes or shifts, which may lead to vanishing gradients or slower training speeds. Batch Normalization (BN) is used to solve this problem. BN smooths the distribution of hidden layer inputs, readjusting the distribution of neuron input values ​​to a standard normal distribution with a mean of 0 and a variance of 1. This places the inputs in a region sensitive to nonlinear functions, so small changes in the input can lead to large differences in the loss function. This solves the vanishing gradient problem present in the sigmoid activation function. Furthermore, larger gradients improve the efficiency of neural network parameter adjustment, resulting in faster convergence.

[0110] The order of nonlinear activation layers and batch normalization (BN) depends on the nonlinear activation function used. Previous studies have consistently used the order "BN + activation function," but in recent years, a few studies have shown that applying BN after the activation function may be more effective. Through experimental analysis, to improve recognition performance, the proposed complete framework ResSENet is built using the GELU + BN order. Experiments show that reversing the GELU + BN order in the network slightly decreases recognition performance, while the average training time to achieve optimal results is longer.

[0111] The downsampling layer is a feature fusion module that embeds the SE attention mechanism: the SE module (Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 7132-7141.) models the correlation between channels by enhancing the learned useful features while suppressing features that do not contribute much to the task. The SE module mainly uses three operations—Squeeze, Excitation, and Reweight—to achieve the above purpose, and its principle is as follows: Figure 8 The Squeeze operation compresses each two-dimensional feature map (channel) into a real number with a global receptive field, so the output dimension corresponds to the number of input channels. Excitation, similar to the "gating" mechanism in Long Short-Term Memory networks, explicitly models channel importance and inter-channel relationships by adjusting the parameter w and using weights. The Reweight operation completes the feature importance labeling along the channel dimension by multiplying the obtained channel weights with the original features channel by channel.

[0112] As network depth increases, problems such as overfitting and vanishing gradients can easily arise. Residual structures, through jump connections, enable effective feature transfer between lower and higher layers, thus avoiding these problems and achieving better generalization ability. The proposed model embeds SE modules into the residual structure to enhance the learning of useful features. Figure 7 As shown, the downsampling layer is a parallel structure that passes the input features through the SE module and a depthwise separable convolution with a kernel size of 3×3 to focus on feature information at different scales. Then, the features obtained from the two branches are residually fused to output richer feature information. Finally, a standard convolution with a stride of 2 and a kernel size of 3×3 is used to halve the feature map size, which facilitates more concise feature extraction in the subsequent process.

[0113] In this embodiment of the invention, a novel load fingerprint construction method is proposed. For the first time, a weighted modal differential field is proposed to encode the current, reactive current, and time-frequency characteristics of the current into three channels of RGB as a load fingerprint for load identification.

[0114] We propose a load identification architecture, ResSENet, based on the design concept of the swin transformer. Compared with existing load identification methods, the model has less training time and a higher recognition rate.

[0115] Figure 9 This is a block diagram illustrating a non-invasive load identification device based on a weighted phantom differential field, according to an exemplary embodiment. The device is used in a non-invasive load identification method based on a weighted phantom differential field. (Refer to...) Figure 9 The device includes an event detection and data preprocessing module 310, an RGB three-channel fusion feature construction module 320, and a load identification model building module 330. Among them:

[0116] The event detection and data preprocessing module 310 is used to acquire a high-frequency public dataset of electricity meter data, perform event detection on the high-frequency public dataset, obtain the current sequence that marks the event transition, decompose and extract the current sequence to obtain the reactive current sequence, and construct the weight of the reactive current sequence based on the harmonic coefficient sequence.

[0117] The RGB three-channel fusion feature construction module 320 is used to obtain a mixed subsequence with rich time-frequency features by performing variational mode decomposition and wavelet packet decomposition on the current sequence, and to construct a two-dimensional RGB three-channel load feature map on the current sequence, the mixed subsequence and the weighted current sequence by using the modal difference field.

[0118] The load identification model building module 330 is used to build a load identification model based on the ResSENet architecture. The two-dimensional load feature map is input into the load identification model to obtain non-intrusive load identification results.

[0119] In this embodiment of the invention, a novel load fingerprint construction method is proposed. For the first time, a weighted modal differential field is proposed to encode the current, reactive current, and time-frequency characteristics of the current into three channels of RGB as a load fingerprint for load identification.

[0120] We propose a load identification architecture, ResSENet, based on the design concept of the swin transformer. Compared with existing load identification methods, the model has less training time and a higher recognition rate.

[0121] Figure 10 This is a schematic diagram of the structure of a non-invasive load identification device provided in an embodiment of the present invention, as shown below. Figure 10 As shown, the non-invasive load identification device may include the above-mentioned Figure 9 The illustrated non-invasive load identification device is based on a weighted modulus differential field. Optionally, the non-invasive load identification device 410 may include a first processor 2001.

[0122] Optionally, the non-intrusive load identification device 410 may also include a memory 2002 and a transceiver 2003.

[0123] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0124] The following is combined with Figure 10 A detailed description of each component of the non-invasive load identification device 410 is provided below:

[0125] The first processor 2001 is the control center of the non-intrusive load identification device 410. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0126] Optionally, the first processor 2001 can perform various functions of the non-intrusive load identification device 410 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0127] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 10 CPU0 and CPU1 are shown in the diagram.

[0128] In a specific implementation, as one example, the non-invasive load identification device 410 may also include multiple processors, for example... Figure 10 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0129] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0130] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently, and may be accessed through the interface circuit of the non-intrusive load identification device 410. Figure 10 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0131] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0132] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 10 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0133] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and can be connected via the interface circuit of the non-intrusive load identification device 410. Figure 10 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0134] It should be noted that, Figure 10 The structure of the non-intrusive load identification device 410 shown does not constitute a limitation on the router. Actual load identification devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0135] Furthermore, the technical effect of the non-invasive load identification device 410 can be referred to the technical effect of the non-invasive load identification method based on weighted modulus difference field described in the above method embodiments, and will not be repeated here.

[0136] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0137] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0138] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0139] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0140] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0141] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0142] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0143] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0144] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0145] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0146] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0147] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A non-intrusive load identification method based on weighted modal difference fields, characterized in that, The method includes: S1. Obtain the high-frequency public dataset of electricity meter data, perform event detection on the high-frequency public dataset, obtain the current sequence that marks the event transition, decompose and extract the current sequence that marks the event transition to obtain the reactive current sequence, and construct the weight of the reactive current sequence based on the harmonic coefficient sequence. The weights for constructing the reactive current sequence based on the harmonic coefficient sequence include: The harmonic coefficient sequence is obtained by fast Fourier transform, and the sum of the first to fifth order harmonic coefficient sequences is used as the weight of the reactive current sequence. S2. By performing variational mode decomposition and wavelet packet decomposition on the current sequence of the marker event transition, a mixed subsequence with rich time-frequency features is obtained. A two-dimensional RGB three-channel load feature map is constructed by weighting the current sequence of the marker event transition, the reactive current sequence, the reactive current sequence, and the mixed subsequence using the modal differential field. The process involves obtaining a mixed subsequence rich in time-frequency features by performing variational mode decomposition and wavelet packet decomposition on the current sequence of the marker event transition, and constructing a two-dimensional RGB three-channel load feature map by weighting the current sequence of the marker event transition, the reactive current sequence, the reactive current sequence, and the mixed subsequence using a modal difference field. Multiple coefficient sets and components are obtained by performing variational mode decomposition and wavelet packet decomposition on the current sequence of the marker event jump. The coefficient sets and components are then pieced together to obtain a hybrid subsequence with rich time-frequency features. The current sequence, reactive current sequence, and mixed subsequence of the marker event transition are reduced in dimensionality using a piecewise aggregation approximation algorithm and the feature space is unified using max-min normalization to obtain the normalized current sequence, reactive current sequence, and mixed subsequence of the marker event transition. The modal differential field is used to transform the normalized current sequence, reactive current sequence and mixed subsequence of the flag event jump into a two-dimensional matrix, which constitutes the RGB three channels of the two-dimensional image. The weight of the reactive current sequence is multiplied by all the values ​​of the reactive current sequence channel to obtain the two-dimensional RGB three-channel load feature map. S3. Build a load identification model based on the ResSENet architecture, input the load feature map into the load identification model, and obtain non-intrusive load identification results.

2. The non-intrusive load identification method based on weighted modulus difference field according to claim 1, characterized in that, The event detection in S1 on the high-frequency public dataset includes: The change in active power is used as the criterion for detecting whether an event transition has occurred to perform event detection on a high-frequency public dataset.

3. The non-invasive load identification method based on weighted modulus difference field according to claim 2, characterized in that, The calculation method for the active power is shown in the following formula (1): （1） In the formula, Indicates active power. This represents the number of data points within a given period. This represents the steady-state current over one cycle. This represents the steady-state voltage over one cycle. The current sequence for the flag event transition is shown in equation (2) below: （2） In the formula, The current sequence that indicates the transition to the flag event. It represents the complete current cycle intercepted from negative to positive at the zero-crossing point of the voltage in a sequence that is in a steady state before the event. It represents the complete current cycle from negative to positive at the zero-crossing point of the voltage in the steady-state sequence after the event.

4. The non-intrusive load identification method based on weighted modulus difference field according to claim 1, characterized in that, The harmonic coefficient sequence in S1 is shown in equation (3) below: （3） In the formula, Represents the harmonic coefficient sequence. , Represents integers, Indicates a time period. Indicates time period Periodic functions within, Represents the imaginary unit. In the real number field The frequency on, Indicates a time step.

5. The non-intrusive load identification method based on weighted modulus difference field according to claim 1, characterized in that, The load identification model in S3 includes: a 3×3 convolutional layer, a first normalized layer, a first convolutional module, a first downsampling module, a second convolutional module, a second downsampling module, a third convolutional module, a third downsampling module, a fourth convolutional module, a global average pooling layer, a second normalized layer, and a linear layer. The convolution module includes a 3×3 convolutional layer, a Gaussian error linear unit, and batch normalization. The downsampling module includes a first branch, a second branch, a 1×1 convolutional layer, and a 3×3 convolutional layer; the first branch includes an average pooling layer, a ReLU-based linear layer, a Sigmoid-based linear layer, and a 1×1 convolutional layer; the second branch includes a 3×3 depthwise separable convolutional layer and a 1×1 convolutional layer.

6. A non-invasive load identification device based on a weighted phantom differential field, wherein the non-invasive load identification device based on a weighted phantom differential field is used to implement the non-invasive load identification method based on a weighted phantom differential field as described in any one of claims 1-5, characterized in that, The device includes: The event detection and data preprocessing module is used to acquire high-frequency public datasets of electricity meter data, perform event detection on the high-frequency public datasets to obtain current sequences that mark event transitions, decompose and extract the current sequences that mark event transitions to obtain reactive current sequences, and construct weights for the reactive current sequences based on the harmonic coefficient sequences. The weights for constructing the reactive current sequence based on the harmonic coefficient sequence include: The harmonic coefficient sequence is obtained by fast Fourier transform, and the sum of the first to fifth order harmonic coefficient sequences is used as the weight of the reactive current sequence. The RGB three-channel fusion feature construction module is used to obtain a mixed subsequence with rich time-frequency features by performing variational mode decomposition and wavelet packet decomposition on the current sequence of the flag event jump. It constructs a two-dimensional RGB three-channel load feature map by weighting the current sequence of the flag event jump, the reactive current sequence, the reactive current sequence, and the mixed subsequence through the modal differential field. The process involves obtaining a mixed subsequence rich in time-frequency features by performing variational mode decomposition and wavelet packet decomposition on the current sequence of the marker event transition, and constructing a two-dimensional RGB three-channel load feature map by weighting the current sequence of the marker event transition, the reactive current sequence, the reactive current sequence, and the mixed subsequence using a modal difference field. Multiple coefficient sets and components are obtained by performing variational mode decomposition and wavelet packet decomposition on the current sequence of the marker event jump. The coefficient sets and components are then pieced together to obtain a hybrid subsequence with rich time-frequency features. The current sequence, reactive current sequence, and mixed subsequence of the marker event transition are reduced in dimensionality using a piecewise aggregation approximation algorithm and the feature space is unified using max-min normalization to obtain the normalized current sequence, reactive current sequence, and mixed subsequence of the marker event transition. The modal differential field is used to transform the normalized current sequence, reactive current sequence and mixed subsequence of the flag event jump into a two-dimensional matrix, which constitutes the RGB three channels of the two-dimensional image. The weight of the reactive current sequence is multiplied by all the values ​​of the reactive current sequence channel to obtain the two-dimensional RGB three-channel load feature map. The load identification model building module is used to build a load identification model based on the ResSENet architecture. It inputs a two-dimensional load feature map into the load identification model to obtain non-intrusive load identification results.

7. A non-invasive load identification device, characterized in that, The non-invasive load identification device includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 5.

Images