Multi-state electric appliance-oriented end-to-end load decomposition method

Abstract

The invention provides an end-to-end load decomposition method for a multi-state electric appliance, and relates to the technical field of load decomposition, and the method comprises the steps: carrying out the data enhancement of a total aggregation power sequence through a time-frequency mixing and random mode, and obtaining a first feature through an encoder and a contrast learning framework model; inputting the first feature and the total aggregation power sequence in the training set into a decoder to obtain the predicted power of the target electric appliance, constructing a time-level hard negative sample feature and an instance-level hard negative sample feature in a contrast learning framework model, and obtaining the predicted power of the target electric appliance based on the two hard negative sample features. According to the method, parameters of an encoder, a decoder and a contrast learning framework model are updated, so that the dependence of the model on the number of labeled samples can be greatly reduced through the construction of hard negative sample features and the encoder expansion convolutional neural network, and the problem of data scarcity is effectively relieved; and meanwhile, the quality and the specificity of feature extraction are remarkably improved, so that the model can capture finer semantic information and a time sequence dependency relationship.

Description

Technical Field

[0001] The present invention relates to the technical field of load decomposition, in particular to an end-to-end load decomposition method for multi-state electrical appliances. Background Art

[0002] In modern society, electricity demand continues to grow, and power system load monitoring has become critical infrastructure for ensuring stable power supply. Traditional intrusive load disaggregation, which collects data from individual appliances through the installation of distributed sensors, enables granular monitoring but suffers from drawbacks such as high hardware deployment costs, complex maintenance, and potential privacy breaches. In contrast, non-intrusive load disaggregation collects overall building energy consumption data through single-point sensing and uses algorithms to infer the operating status and energy consumption distribution of individual appliances, significantly reducing system costs and improving usability.

[0003] However, with the widespread adoption of smart homes, industrial automation upgrades, and growing electricity demand across diverse scenarios, the number and variety of multi-state appliances has exploded. The complex dynamic characteristics of these appliances (frequent power fluctuations, multiple operating modes, and nonlinear transient responses) make accurate extraction of their operating states difficult, creating a core bottleneck hindering the accuracy of non-intrusive load decomposition. This presents challenges, including difficulty in feature extraction, signal aliasing caused by multiple appliances operating, and complex data distribution.

[0004] In non-intrusive load decomposition, traditional signal processing and statistical analysis methods often face the problems of insufficient feature expression ability and poor generalization when processing the above-mentioned complex time series data. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention aims to propose an end-to-end load decomposition method for multi-state electrical appliances, comprising: Step 1: Collect the total power of all appliances in a certain period of time through smart meters and obtain the total aggregate power sequence Y=[y1,y2,…,y T ], where y1 is the instantaneous value of the total power at time 1, y2 is the instantaneous value of the total power at time 2, and y T is the instantaneous value of the total power at time T; at the same time, the real power of the target appliance corresponding to each moment is obtained; Step 2: Divide the total aggregate power sequence by sliding the window to obtain multiple total power subsequences ,in, represents the first total power subsequence, represents the second total power subsequence, represents the Nth total power subsequence; Step 3: For each total power sub-sequence, perform time-frequency mixing and random cropping data enhancement operations on the total power sub-sequence to obtain a time-frequency mixed sequence and a cropped sequence; Step 4: Input the total power sub-sequence, the time-frequency mixed sequence corresponding to the total power sub-sequence, and the cropped sequence into the encoder to obtain a high-dimensional feature representation; Step 5: Input the high-dimensional feature representation into the contrast learning framework model to obtain a first feature, and input the first feature and the total aggregated power sequence Y in the training set into the decoder to obtain a predicted power of the target appliance; Step 6: Based on the high-dimensional feature representation, construct a time-level hard negative sample feature and an instance-level hard negative sample feature, and then calculate a loss generated by the hard negative sample feature ; Step 7: Based on the true power and the predicted power of the target appliance, calculate a reconstruction loss ; Step 8: Weighted sum of the loss generated by the hard negative sample feature and the reconstruction loss to obtain a total loss , and based on the total loss , iteratively update the parameters in the encoder, the parameters in the contrast learning framework model, and the parameters in the decoder to obtain a trained encoder, a trained contrast learning framework model, and a trained decoder; Step 9: Based on the trained encoder, the trained contrast learning framework model, and the trained decoder, perform steps 2 to 5 on the total aggregated power sequence in the test set to obtain the power of the target appliance.

[0006] Optionally, step 3 specifically includes: Step 3.1: For each total power sub-sequence, perform time-frequency mixing on each total power sub-sequence to obtain a time-frequency mixed sequence; Specifically, perform a fast Fourier transform on the total power sub-sequence to obtain a frequency domain sequence, replace the frequency components of the frequency domain sequence with the frequency components of another total power sub-sequence to obtain a mixed frequency domain sequence, and perform an inverse fast Fourier transform on the mixed frequency domain sequence to obtain a time-frequency mixed sequence; Step 3.2: Randomly crop each total power sub-sequence to obtain a cropped sequence; Specifically, randomly select two time-overlapping time periods from the time period of the total power sub-sequence, and crop the sequences in the overlapping part of the two time periods to obtain a cropped sequence.

[0007] Optionally, the temporal hard negative sample feature in step 6 is constructed using the following formula: ; ; in, and is the total power subsequence Y at timestamp t i Temporal-level hard negative features; is the randomly selected mixing coefficient for the anchor point, ; and is the total power subsequence Y at timestamp t i The corresponding high-dimensional feature representation tensor, and For timestamp Total power subsequence The corresponding high-dimensional feature representation tensor, ∈ , A collection of timestamps that overlap when randomly cropped.

[0008] Optionally, the instance-level hard negative sample feature in step 6 is constructed using the following formula: ; ; in, and is the total power subsequence Y at timestamp t i Instance-level hard negative sample features; is the randomly selected mixing coefficient for the anchor point, ; and is the total power subsequence Y at timestamp t i The corresponding high-dimensional feature representation tensor, and is the total power subsequence Y at timestamp t j The corresponding high-dimensional feature representation tensor, where the total power subsequence Y j For all total power subsequences except the total power subsequence Y i Other than the sequence.

[0009] Optionally, calculate the loss of hard negative features in step 6 , specifically including: Step A1: Calculate the temporal loss based on the temporal-level hard negative features and the instance-level hard negative features and instance loss , which is specifically achieved through the following formula: ; ; in, , , , express The set of timestamps other than t in , , express In addition to the total power subsequence Y i The set of total power subsequences other than , where is the total number of total power subsequences, , ; Step A2: Based on time loss and instance loss , calculate the loss generated by the hard negative sample feature , which is specifically achieved through the following formula: ; Wherein, T is the maximum value of the timestamp of the total power subsequence.

[0010] Optionally, step 7 is specifically implemented by the following formula: ; in, Represents the total power subsequence Y i The observation mask of is the total power subsequence Y i The corresponding predicted power, is the total power subsequence Y i The corresponding real power, is the sequence obtained by mixing the time-frequency mixed sequence corresponding to the total power subsequence and the pruned sequence, for The predicted power, for The observation mask.

[0011] Optionally, the loss of hard negative features in step 8 and MSE loss Weighted sum to get the total loss , which is specifically achieved through the following formula: ; in, is a hyperparameter.

[0012] The beneficial effects of adopting the above technical solution are: The present invention performs data enhancement on the total aggregate power sequence through time-frequency mixing and randomization, and then extracts high-dimensional feature representations through the encoder. The high-dimensional feature representations are input into the contrastive learning framework model, and the obtained first feature and the total aggregate power sequence in the training set are input into the decoder to obtain the predicted power of the target electrical appliance. At the same time, hard negative sample features at the time level and hard negative sample features at the instance level are constructed. Based on the two hard negative sample features, the parameters of the encoder, decoder and contrastive learning framework model are updated. Through the construction of hard negative sample features and the synergistic effect of the encoder's expanded convolutional neural network, the model's dependence on the number of labeled samples is greatly reduced, effectively alleviating the problem of data scarcity; at the same time, the quality and specificity of feature extraction are significantly improved, enabling the model to capture more detailed semantic information and temporal dependencies. While taking into account computational efficiency, this method breaks through the performance limitations of traditional methods and improves the decomposition performance of multi-state electrical appliances. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Schematic diagram of the structure of an end-to-end load decomposition method for multi-state electrical appliances in an embodiment of the present invention; Figure 2 4 is a flow chart of an end-to-end load decomposition method for multi-state electrical appliances in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0015] In response to the problems existing in the prior art, the present invention provides an end-to-end load decomposition method for multi-state electrical appliances, aiming to overcome the problem of difficult feature extraction caused by the complex dynamic characteristics of multi-state electrical appliances. A data enhancement strategy that includes time series awareness and a hard negative sample generation mechanism based on time and instance dual benchmarks are designed. By generating negative samples with more discriminative value (i.e., samples that are similar to the anchor point features but have different actual categories), the model's ability to distinguish similar electrical appliance states is effectively enhanced. By introducing new data enhancement technology and a contrastive learning framework, the accuracy and robustness of non-invasive load decomposition in multi-state electrical appliance decomposition are significantly improved by generating high-quality training samples and strengthening feature discrimination capabilities. Experiments on two public data sets show that this method significantly improves the decomposition accuracy of each electrical appliance compared with traditional solutions, especially for multi-state electrical appliances, and the decomposition accuracy is significantly improved compared to other algorithms.

[0016] Specifically, the application proposes an end-to-end contrastive learning framework. Unlike traditional two-stage contrastive learning methods, this method does not require a complex fine-tuning process. Instead, a certain amount of data of the target electrical appliance is directly used as prior information to the input end and is synchronized with the total power data to enter the model processing flow. In this process, an expanded convolutional neural network is designed to extract features from the input sequence data. The unique dilation coefficient setting of the model enables it to flexibly adapt to the timing characteristics of different electrical appliances and effectively capture long and short time dependencies. Subsequently, these features that have incorporated the prior information of the target electrical appliance directly enter the contrastive learning framework. Through the design of a hard negative sample construction mechanism, more discriminative negative sample pairs are generated, and a contrastive loss function is used for learning and optimization. Finally, the model can accurately decouple the power consumption of the target electrical appliance from the total power signal of the test set, realizing end-to-end integrated processing from data input to result output, significantly improving the efficiency and accuracy of load decomposition, and avoiding the complex post-processing and independent fine-tuning steps in traditional methods.

[0017] The non-intrusive load decomposition problem can be regarded as a regression problem and can be defined as a method for estimating the energy consumption of a single electrical appliance using only the total aggregated power data measured at the electrical inlet of a building. Mathematically, the decomposition problem can be formulated as follows: ; The non-intrusive load decomposition algorithm decomposes the total aggregated power data into the power consumed by each electrical appliance, minimizing the reconstruction error. To achieve this goal, it is necessary to collect aggregated data such as voltage, current, and power at the electrical inlet of a building and process it on a digital processor using complex software. These operations are usually performed using a smart meter.

[0018] In combination Figure 1 , the end-to-end load decomposition method for multi-state electrical appliances provided by the application is composed of three key modules: 1) Data augmentation module: total aggregated power and electrical appliance power are augmented through time-frequency mixing and random methods; 2) Contrastive feature learning module: the data described above are extracted by an encoder to further construct hard negative sample pairs in the feature space. Using the designed hard negative sample selection strategy, the similarity of similar features is maximized while the discriminative ability between classes is enhanced, thereby achieving more discriminative feature representation learning. 3) Power decoupling module: the learned high-dimensional features are input into the decoder along with the total power in the test phase, and the time series power estimate of the target electrical appliance is output.

[0019] Specifically, in combination Figure 1 and Figure 2 , the following steps can be included: Step 1: Collect the total power of all electrical appliances for a certain period of time through a smart meter to obtain a total aggregated power sequence Y = [y1, y2, …, yT], where T represents the total number of time points.T ], where y1 is the instantaneous value of the total power at time 1, y2 is the instantaneous value of the total power at time 2, and y T is the instantaneous value of the total power at time T; at the same time, the real power of the target appliance corresponding to each moment is obtained , these data are all time series data.

[0020] Since aggregated power data is often long and continuous, directly processing the entire segment of data may lead to excessive consumption of computing resources and also affect the learning efficiency of the model. Therefore, a sliding window method is used to process the data in segments.

[0021] The core idea of ​​the sliding window method is to divide the continuous aggregate power data into multiple segments of fixed length, each segment is represented as a window, and the length of the window is represented by N, N=1024.

[0022] Step 2: Divide the total aggregate power sequence by sliding the window to obtain multiple total power subsequences ,in, represents the first total power subsequence, represents the second total power subsequence, represents the Nth total power subsequence; Step 3: For each total power subsequence, perform data augmentation operations of time-frequency mixing and random cropping on the total power subsequence to obtain the time-frequency mixed sequence and the cropped sequence; Compared to traditional enhancement methods in contrastive learning (such as noise addition and time scaling), this solution uses time-frequency mixing and random cropping to enhance data before feature extraction. When processing time series data, using time-frequency mixing can strengthen periodic features, better capturing data characteristics over longer periods of time while weakening random fluctuations. Random cropping, on the other hand, preserves local temporal characteristics in the original data and is suitable for capturing characteristics over shorter periods of time.

[0023] Step 3.1: For each total power subsequence, perform time-frequency mixing on each total power subsequence to obtain a time-frequency mixed sequence; Specifically, a fast Fourier transform is performed on the total power subsequence to obtain a frequency domain sequence, the frequency components of the frequency domain sequence are replaced with the frequency components of the frequency domain sequence of another total power subsequence to obtain a mixed frequency domain sequence, and the mixed frequency domain sequence is inversely fast Fourier transformed to obtain a time-frequency mixed sequence; Among them, exchanging frequency components between frequency domain sequences does not introduce additional noise and can provide more reliable enhancement for preserving the semantic characteristics of the data.

[0024] Step 3.2: Randomly crop each total power subsequence to obtain a cropped sequence; Specifically, two overlapping time periods are randomly selected from the time period of the total power subsequence, the sequences of the overlapping parts of the two time periods are clipped, and the clipped sequences of the overlapping parts of the two time periods are used as the clipped sequences.

[0025] For example, one can randomly sample two overlapping time periods and ,in, The overlapping part of the two time periods is , and then cut The sequence is taken as the pruned sequence.

[0026] Step 4: Input the total power subsequence, the time-frequency mixed sequence corresponding to the total power subsequence, and the cropped sequence into the encoder to obtain a high-dimensional feature representation; Among them, the encoder is mainly composed of an expanded convolutional neural network, which significantly expands the receptive field through the design of the expansion factor without increasing the number of parameters and computational complexity. It can capture a wider range of contextual information, is more suitable for processing long sequence data, and can efficiently model long-distance time dependencies with lower computational complexity.

[0027] The encoder is Figure 1 The Encoder fΦ in , the high-dimensional features include Figure 1 Z in 1:M+K 、 and Z m .

[0028] Step 5: Input the high-dimensional feature representation into the contrastive learning framework model to obtain the first feature. Input the first feature and the total aggregate power sequence Y in the training set into the decoder to obtain the predicted power of the target appliance. Among them, the first feature includes Figure 1 in , the decoder is Figure 1 Decoder gθ in .

[0029] Because traditional contrastive learning feature spaces contain a large number of easy negative examples and a small number of hard negative examples, easy negative examples are those that are far from the anchor point (i.e., positive examples) and contribute less to the contrastive loss. To enhance the model's ability to discriminate complex features, a dual hard negative feature construction strategy is adopted. Hard negative features are constructed from both instance and time dimensions. These constructed hard negative features are then used to further optimize the extracted features during contrastive learning. Hard negative features are trained and continuously optimized alongside high-dimensional features. Driven by the loss function, a gradient signal is generated, guiding the model to dynamically adjust the feature representation to obtain more accurate features. For detailed implementation, refer to steps 6 and 7.

[0030] Step 6: Based on the high-dimensional feature representation, construct the time-level hard negative sample features and instance-level hard negative sample features, and then calculate the loss generated by the hard negative sample features ; The time-level hard negative sample feature is constructed by the following formula: ; ; in, and is the total power subsequence Y at timestamp t i Temporal-level hard negative features; is the randomly selected mixing coefficient for the anchor point, ; and is the total power subsequence Y at timestamp t i The corresponding high-dimensional feature representation tensor, and For timestamp Total power subsequence The corresponding high-dimensional feature representation tensor, ∈ , A collection of timestamps that overlap when randomly cropped.

[0031] The instance-level hard negative sample feature is constructed using the following formula: ; ; in, and is the total power subsequence Y at timestamp t i Instance-level hard negative sample features; is the randomly selected mixing coefficient for the anchor point, ; and a total power sub-sequence Y i a tensor of corresponding high-dimensional feature representations, and a total power sub-sequence Y j a tensor of corresponding high-dimensional feature representations, wherein the total power sub-sequence Y j is a sequence other than the total power sub-sequence Y i .

[0032] wherein a loss generated by the hard negative sample feature is calculated , specifically comprising: Step A1: based on the time-level hard negative sample feature and the instance-level hard negative sample feature, calculating a time loss and an instance loss , specifically implemented by the following formula: ; ; wherein, , , , denotes a set of other time stamps in other than t, wherein, is a total number of total power sub-sequences, denotes a set of total power sub-sequences in other than the total power sub-sequence Y i . ; Step A2: according to the time loss and the instance loss , calculating a loss generated by the hard negative sample feature , specifically implemented by the following formula: ; wherein T is a maximum value of time stamps of total power sub-sequences.

[0033] Step 7: based on the real power and the predicted power of the target electrical appliance, calculating a reconstruction loss , specifically implemented by the following formula: ; wherein, denotes an observation mask of the total power sub-sequence Y i , is a corresponding predicted power of the total power sub-sequence Y i , is a corresponding predicted power of the total power sub-sequence Y i ​corresponding real power, a sequence obtained by mixing the sequence after time-frequency mixing and the sequence after clipping corresponding to the total power sub-sequence, a predicted power, an observation mask. a predicted power, an observation mask.

[0034] Step 8: loss generated by hard negative sample features and reconstruction loss weighted sum to obtain total loss , which is specifically realized by the following formula: ; wherein, is a hyperparameter.

[0035] and based on the total loss the parameters in the encoder, the parameters in the contrastive learning framework model and the parameters in the decoder are iteratively updated to obtain the trained encoder, the trained contrastive learning framework model and the trained decoder; Step 9: based on the trained encoder, the trained contrastive learning framework model and the trained decoder, the total aggregated power sequence in the test set is decomposed according to steps 2 to 5 to obtain the power corresponding to the target appliance.

[0036] That is, by sliding window, the total aggregated power sequence in the test set is divided to obtain a plurality of total power sub-sequences, and the total power sub-sequences are time-frequency mixed and randomly clipped to obtain a sequence after time-frequency mixing and a sequence after clipping. The total power sub-sequences, the sequence after time-frequency mixing and the sequence after clipping corresponding to the total power sub-sequences are input into the trained encoder to obtain high-dimensional feature representation, the high-dimensional feature representation is input into the trained contrastive learning framework model to obtain the first feature, and the first feature and the total aggregated power sequence Y in the test set are input into the trained decoder to obtain the power corresponding to the target appliance.

[0037] The key point of the present application is: Construction of double hard negative samples: since the traditional negative sample includes a large number of easy negative samples and a small number of difficult negative samples, the distance between the easy negative sample and the anchor point is too far to contribute to the contrast loss. In real life, the states of appliances are very similar and not easy to distinguish. In order to better distinguish the states of appliances, the present application proposes to construct hard negative samples, i.e. hard negative samples with higher similarity to anchor points, so that the features of similar appliances can be more clearly distinguished.

[0038] An end-to-end load decomposition algorithm based on contrastive learning: In view of the fact that traditional supervised methods require a large number of labels, there are bottlenecks of high data acquisition cost and low labeling efficiency, and the features extracted by conventional contrastive learning methods have limited generalization ability, which is difficult to meet the requirements of complex scenes. The scheme innovatively embeds the optimized contrastive learning mechanism into the encoder-decoder architecture, and through the synergistic effect of double hard negative sample construction and dilated convolutional neural network, on the one hand, the dependence of the model on the number of labeled samples is greatly reduced, effectively alleviating the problem of data scarcity; on the other hand, the quality and specificity of feature extraction are significantly improved, so that the model can capture more fine semantic information and temporal dependence. While taking into account the computational efficiency, this method breaks through the performance limitations of traditional methods and improves the decomposition performance of multi-state appliances.

[0039] Based on the above technical solutions, the proposed end-to-end method based on contrastive learning is compared with two types of mainstream benchmark methods in the task of multi-state appliance load decomposition: traditional contrastive learning method (unsupervised learning) and full supervised learning method. The experimental results show that under the premise of consistent experimental conditions, the proposed method is significantly better than the contrast algorithm in terms of overall performance: not only does the overall decomposition accuracy improve dramatically, but it also shows significant advantages in the complex scene of multi-state appliances. The specific results are as follows: Table 1 Results on the uk-dale dataset

[0040] According to the experimental comparison results in Table 1, compared with traditional supervised algorithms and typical representation learning models, the improved method proposed in this paper has achieved significant improvement in the decomposition task of multi-state and single-state appliances. Through multi-dimensional performance evaluation indicators, it can be seen that: (1) In the multi-state appliance decomposition task, the MAE (Mean Absolute Error) index is improved by more than 55%, indicating that the algorithm's error suppression ability is significantly enhanced; (2) The EpD (Energy Imbalance Degree) index is also optimized by nearly 50% in multi-state appliances, verifying the improvement of the algorithm's energy distribution fitting accuracy; (3) At the same time, SAE (Average Symbol Error) also shows a positive optimization trend, further verifying the model's ability to capture temporal dynamic processes.

[0041] (4) For the decomposition task of single-state appliances, the improved method also shows superior performance in the above three evaluation indicators.

[0042] The above description is merely an illustration of the preferred embodiments of the present disclosure and the technical principles employed. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present disclosure is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present disclosure.

Claims

1. An end-to-end load decomposition method for multi-state electrical appliances, characterized in that: include: Step 1: Collect the total power of all appliances in a certain period of time through smart meters and obtain the total aggregate power sequence Y=[y1,y2,…,y T ], where y1 is the instantaneous value of the total power at time 1, y2 is the instantaneous value of the total power at time 2, and y T is the instantaneous value of the total power at time T; at the same time, the real power of the target appliance corresponding to each moment is obtained; Step 2: Divide the total aggregate power sequence through a sliding window to obtain multiple total power subsequences ,in, represents the first total power subsequence, represents the second total power subsequence, represents the Nth total power subsequence; Step 3: For each total power subsequence, perform data augmentation operations of time-frequency mixing and random cropping on the total power subsequence to obtain the time-frequency mixed sequence and the cropped sequence; Step 4: Input the total power subsequence, the time-frequency mixed sequence corresponding to the total power subsequence, and the cropped sequence into the encoder to obtain a high-dimensional feature representation; Step 5: Input the high-dimensional feature representation into the contrastive learning framework model to obtain the first feature. Input the first feature and the total aggregate power sequence Y in the training set into the decoder to obtain the predicted power of the target appliance. Step 6: Based on the high-dimensional feature representation, construct the time-level hard negative sample features and instance-level hard negative sample features, and then calculate the loss generated by the hard negative sample features ; Step 7: Calculate the reconstruction loss based on the actual power and predicted power of the target appliance ; Step 8: Loss on hard negative features and reconstruction loss Weighted sum to get the total loss , and based on the total loss Iteratively updating the parameters in the encoder, the parameters in the contrastive learning framework model, and the parameters in the decoder to obtain a trained encoder, a trained contrastive learning framework model, and a trained decoder; Step 9: Based on the trained encoder, the trained contrastive learning framework model, and the trained decoder, follow steps 2 to 5 to decompose the total aggregate power sequence in the test set to obtain the power of the target appliance.

2. The end-to-end load decomposition method for multi-state electrical appliances according to claim 1, characterized in that: Step 3 specifically includes: Step 3.1: For each total power subsequence, perform time-frequency mixing on each total power subsequence to obtain a time-frequency mixed sequence; Specifically, a fast Fourier transform is performed on the total power subsequence to obtain a frequency domain sequence, the frequency components of the frequency domain sequence are replaced with the frequency components of the frequency domain sequence of another total power subsequence to obtain a mixed frequency domain sequence, and the mixed frequency domain sequence is inversely fast Fourier transformed to obtain a time-frequency mixed sequence; Step 3.2: Randomly crop each total power subsequence to obtain a cropped sequence; Specifically, two overlapping time periods are randomly selected from the time period of the total power subsequence, the sequences of the overlapping parts of the two time periods are clipped, and the clipped sequences of the overlapping parts of the two time periods are used as the clipped sequences.

3. The end-to-end load decomposition method for multi-state electrical appliances according to claim 2, characterized in that: The temporal hard negative feature described in step 6 is constructed using the following formula: ； ； in, and is the total power subsequence Y at timestamp t i Temporal-level hard negative features; is the randomly selected mixing coefficient for the anchor point, ; and is the total power subsequence Y at timestamp t i The corresponding high-dimensional feature representation tensor, and For timestamp Total power subsequence The corresponding high-dimensional feature representation tensor, ∈ , A collection of timestamps that overlap when randomly cropped.

4. The end-to-end load decomposition method for multi-state electrical appliances according to claim 3, characterized in that: The instance-level hard negative sample features described in step 6 are constructed using the following formula: ； ； in, and is the total power subsequence Y at timestamp t i Instance-level hard negative sample features; is the randomly selected mixing coefficient for the anchor point, ; and is the total power subsequence Y at timestamp t i The corresponding high-dimensional feature representation tensor, and is the total power subsequence Y at timestamp t j The corresponding high-dimensional feature representation tensor, where the total power subsequence Y j For all total power subsequences except the total power subsequence Y i Other than the sequence.

5. The end-to-end load decomposition method for multi-state electrical appliances according to claim 4, characterized in that: The loss generated by calculating the hard negative sample features in step 6 , specifically including: Step A1: Calculate the temporal loss based on the temporal-level hard negative features and the instance-level hard negative features and instance loss , which is specifically achieved through the following formula: ； ； in, , , , express The set of timestamps other than t in , , express In addition to the total power subsequence Y i The set of total power subsequences other than , where is the total number of total power subsequences, , ; Step A2: Based on time loss and instance loss , calculate the loss generated by the hard negative sample feature , which is specifically achieved through the following formula: ； Wherein, T is the maximum value of the timestamp of the total power subsequence.

6. The end-to-end load decomposition method for multi-state electrical appliances according to claim 1, characterized in that: Step 7 is specifically implemented by the following formula: ； in, Represents the total power subsequence Y i The observation mask is the total power subsequence Y i The corresponding predicted power, is the total power subsequence Y i The corresponding real power, is the sequence obtained by mixing the time-frequency mixed sequence corresponding to the total power subsequence and the pruned sequence, for The predicted power, for The observation mask.

7. The end-to-end load decomposition method for multi-state electrical appliances according to claim 1, characterized in that: The loss of hard negative features in step 8 The total loss is obtained by weighted summation of the MSE loss, which is implemented by the following formula: ； in, is a hyperparameter.

Images