Electric appliance classification method and data processing equipment

Abstract

The invention provides an electric appliance classification method. The electric appliance classification method specifically comprises a data acquisition step, a target electric appliance identification step, a feature extraction step and a target electric appliance classification step. According to the electric appliance classification method, decoupling of a complex scene where target electric appliances are overlapped is achieved through the MICN, and accurate recognition of different types of target electric appliances is achieved. According to the method, the target electric appliance can be accurately identified and classified by combining with the training feature extractor, so that the accuracy and adaptability of secondary classification of the target electric appliance are improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and in particular to an electrical appliance classification method and data processing equipment. Background Art

[0002] In the field of power monitoring, power load monitoring technologies are mainly divided into two types: intrusive load monitoring (ILM) and non-intrusive load monitoring (NILM). Intrusive load monitoring technology embeds sensors in each electrical appliance to collect equipment operating parameters in real time to achieve accurate monitoring, but it is costly, complex to install and maintain, and may have an adverse effect on equipment performance and life. In contrast, non-intrusive load monitoring technology analyzes the overall power consumption data of smart meters, extracts features with the help of algorithms, and splits the power consumption behavior of electrical appliances without modifying equipment or installing additional sensors. This method is low-cost, easy to deploy, and has good adaptability and scalability. It provides important technical support for large-scale power consumption monitoring and energy-saving management, but the existing non-intrusive load decomposition algorithms still face many challenges.

[0003] There are usually multiple target electrical appliances under the total power supply. Target electrical appliances account for an important proportion of residential electricity consumption, and their working modes are diverse. Different types of electrical appliances have significant differences in power consumption, working hours, waveform characteristics, etc. Accurate identification and classification of these target electrical appliances are crucial to achieve refined perception of residential electricity consumption behavior. On the one hand, the perception of residential electricity consumption behavior can provide users with detailed electricity consumption information, help users optimize electricity consumption habits, reduce energy consumption, and achieve energy conservation and emission reduction goals. For example, by understanding the frequency and duration of use of different target electrical appliances, users can reasonably arrange the use time of electrical appliances to avoid concentrated electricity consumption during peak hours. On the other hand, for power suppliers, accurate perception of electricity consumption behavior helps optimize grid load scheduling, improve grid operation efficiency, and ensure the stability and reliability of power supply.

[0004] At present, the existing technology still has shortcomings in the secondary classification of target electrical appliances. Most methods have achieved certain results in the primary classification. The primary classification of electrical appliances refers to the classification of electrical appliances into heating and non-heating categories. However, the accuracy and adaptability of the secondary classification of electrical appliances are not high. The secondary classification of electrical appliances refers to the further subdivision of different types of target electrical appliances, such as electric kettles, rice cookers, water heaters, etc. When processing complex power waveforms of target electrical appliances, the existing technology has difficulty in accurately capturing its characteristic changes, resulting in classification errors or failure to identify new types of electrical appliances. In addition, the electricity usage habits of different household users vary greatly, and the usage scenarios of electrical appliances are complex and diverse. Traditional methods are difficult to adapt to these changes and cannot meet the needs of actual applications. Summary of the invention

[0005] The present invention provides an electrical appliance classification method and data processing equipment to solve the technical problems of poor accuracy and adaptability in solving the electrical appliance secondary classification problem in the prior art.

[0006] In order to solve the above problems, the present invention provides an electrical appliance classification method, which specifically includes the following steps. A data collection step, collecting a voltage sequence U and a current sequence I of a total power supply at a certain frequency, and obtaining the active power sequence P and the reactive power sequence Q of the total power supply. A target appliance identification step, obtaining the information of each target appliance under the total power supply through the active power sequence P and the reactive power sequence Q of the total power supply, and using the MICN model to process the information of each target appliance, decoupling and splitting the coupling relationship between different appliances, so as to obtain the characteristics of a single target appliance. A feature extraction step, constructing a feature extractor according to the historically accumulated load data characteristics of each target appliance, and using the feature extractor to obtain the feature representation of a single target appliance. A target appliance classification step, taking the feature representation of the single target appliance obtained in the feature extraction step and the static input as input, and realizing the prediction of the secondary category of the target appliance through the TFT-C model.

[0007] Furthermore, the target electrical appliance identification step specifically includes the following steps: a target electrical appliance event extraction step, obtaining each target electrical appliance h under the total power supply according to the active power sequence P and the reactive power sequence Q id information, the target appliance h id The information includes the opening time Closing time Active power sequence during operation And the reactive power sequence during operation The first target appliance clustering step is to cluster the target appliances h at a certain time interval. id Filter out the target appliances that are naturally turned off; cluster the target appliances that are naturally turned off so that the time interval between every two adjacent target appliances is greater than a preset time, so as to obtain K groups of target appliances The second target appliance clustering step uses the MICN model to cluster each group of target appliances. Further clustering is performed to finally obtain M clusters of target appliances C M , the target electrical appliance C M The information includes the first opening time Last closing time Active power sequence Reactive power sequence

[0008] Furthermore, in the second target electrical appliance clustering step, the MICN model is used to cluster each group of target electrical appliances. The step of further clustering specifically includes the following steps:

[0009] Artificially construct the target electrical appliance superposition use case, use the superimposed use case as the model input, use the model output and the target electrical appliance waveform before superposition to calculate the model loss, train the neural network, and build the MICN model. Extract the high-frequency details of the appliance, and aggregate the time steps through the Global module of the MICN model to capture the long-term operation mode of the appliance. Through the mixing of high-frequency and low-frequency information, identify the unique operation characteristics of different appliances, decouple the waveforms of the target electrical appliances running at the same time, and obtain a pure single-type target electrical appliance waveform.

[0010] Furthermore, in the feature extraction step, a feature extractor is constructed according to the historical accumulated load data features of each target electrical appliance. This step specifically includes the following steps: obtaining waveform data of a certain frequency of each type of target electrical appliance as a positive sample, and obtaining waveform data of a certain frequency of target electrical appliances other than this type as a negative sample; inputting the positive sample and the negative sample into the BI-LSTM autoencoder, compressing them into a low-dimensional representation and reconstructing them, using the low-dimensional output of the encoder as the feature of the sample, calculating the contrast loss between the positive sample pair and the negative sample pair, optimizing the sample reconstruction loss and the loss between the positive and negative samples to optimize the BI-LSTM encoder parameters, so that the encoder can become a feature extractor.

[0011] Furthermore, the loss function formula of the BI-LSTM encoder is as follows:

[0012] X={X + ,X -}

[0013] X′={X′ + ,X′ -}

[0014] D′={D′ + ,D′ -}

[0015] Loss = mse(X,X′)+sim(D′) i,+ ,D′ j,+ )-sim(D′ i,+ ,D′ k,- )

[0016] Among them, X is the model input, X + represents positive samples, X- represents negative samples, X′ is the output of the autoencoder, D′ is the feature vector after partial compression of the encoder, mse represents mean square error, and sim represents cosine similarity.

[0017] Furthermore, in the target electrical appliance classification step, the calculation formula of the TFT-C model is as follows:

[0018]

[0019]

[0020] T″=Patch(T′)

[0021] O=Self-attention(T″,S)

[0022]

[0023] Output = Softmax(O′)

[0024] in, is the time when the target appliance is turned on, Variable_select() is the variable selection network, is the active power sequence of the target electrical appliance, is the reactive power sequence of the target appliance, Lstm_encoder() is the LSTM encoder, GRN() is the gated residual network, Patch() and Self-attention() are attention mechanisms used to divide the time series data of a certain frequency, MLP() is a multi-layer perceptron, Softmax() is a normalized exponential function, and Output represents the secondary category of the target appliance that is finally output.

[0025] Furthermore, in the target electrical appliance classification step, the static input includes the season when the target electrical appliance is operating, the temperature when the target electrical appliance is operating, and the time when the target electrical appliance starts operating.

[0026] Furthermore, after the target electrical appliance classification step, the electrical appliance classification method further includes an encoder updating step, taking the prediction of the target electrical appliance secondary category output by the target electrical appliance classification step as input, adjusting the parameters of the BI-LSTM encoder, and only considering the reconstruction loss of the sample, the loss function formula is as follows:

[0027] Loss = mse(X,X′)

[0028] Where mse represents mean square error, X is the model input, and X′ is the autoencoder output.

[0029] Furthermore, after the encoder updating step, the electrical appliance classification method further includes a reclassification step, using the BI-LSTM encoder adjusted in the encoder updating step to re-match the target electrical appliance that was not successfully classified in the target electrical appliance classification step to determine the target electrical appliance category to which it belongs.

[0030] The present invention also includes a data processing device, the data processing device includes a memory, the memory is used to store executable program code. The data processing device also includes a processor, which is used to read the executable program code to run a computer program corresponding to the executable program code to perform the steps in the above-mentioned electrical appliance classification method.

[0031] The advantage of the present invention is that the present invention provides an electrical appliance classification method, which uses the MICN network to achieve decoupling of complex scenes superimposed on target electrical appliances, achieves accurate identification of different types of target electrical appliances, and greatly simplifies the difficulty of subsequent classification of different types of target electrical appliances. The present invention can accurately identify and classify target electrical appliances by combining training feature extractors, with higher accuracy than traditional methods. The present invention can also fine-tune the parameters of the feature extractor through the user's target electrical appliance data, achieve adaptive updating of algorithm parameters, and construct a feature extraction scheme that fits user habits to adapt to the differences in electricity usage habits of different household users and complex and diverse electrical appliance usage scenarios, so as to improve the accuracy and adaptability of the secondary classification of target electrical appliances. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flow chart of an electrical appliance classification method according to an embodiment of the present invention;

[0033] Figure 2 is a flow chart of target electrical appliance identification steps in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the MICN structure in an embodiment of the present invention;

[0035] Figure 4 Schematic diagram of the structure of the TFT-C model in an embodiment of the present invention. Specific embodiments

[0036] The following describes the preferred embodiments of the present invention with reference to the drawings in the specification to illustrate that the present invention can be implemented. These embodiments can fully introduce the technical content of the present invention to those skilled in the art, making the technical content of the present invention clearer and easier to understand. However, the present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0037] like Figure 1As shown, this embodiment provides an electrical appliance classification method, and the electrical appliance classification method specifically includes steps S1 to S6.

[0038] Step S1: data collection step, collecting a voltage sequence U and a current sequence I of a total power supply at a frequency of 6400 Hz, and obtaining an active power sequence P and a reactive power sequence Q of the total power supply.

[0039] P=UI cosφ

[0040] Q=UIsinφ

[0041] Among them, U represents voltage, I represents current, φ represents the angle between voltage and current, P represents active power, and Q represents reactive power.

[0042] Step S2: target appliance identification step, obtaining information of each target appliance under the total power supply through the active power sequence P and reactive power sequence Q of the total power supply, and using the MICN model to process the information of each target appliance, decoupling and splitting the coupling relationship between different appliances to obtain the characteristics of a single target appliance.

[0043] like Figure 2 As shown, the target electrical appliance identification step specifically includes steps S21 to S23.

[0044] Step S21: Target electrical appliance event extraction step, obtaining each target electrical appliance h under the total power supply according to the active power sequence P and the reactive power sequence Q id information, the target appliance h id The information includes the opening time Closing time Active power sequence during operation And the reactive power sequence during operation

[0045] Step S22: First target electrical appliance clustering step, from the target electrical appliances h id Filter out the target appliances that are naturally turned off; cluster the target appliances that are naturally turned off so that the time interval between every two adjacent target appliances is greater than 20 minutes, and consider that a new cluster of target appliances is turned on to obtain K groups of target appliances

[0046] Step S23: The second target appliance clustering step uses the MICN model to cluster each group of target appliances. Further clustering is performed.

[0047] like Figure 3As shown in Figure 1, the Multi-Scale Equidistant Convolutional Network (MICN) uses multiple branches with different convolution kernels to model different potential patterns of the sequence. For each branch, a local module based on downsampling convolution is used to extract local features of the sequence, and on this basis, a global module based on equidistant convolution is used to model global correlation. This design reduces the time and space complexity to linear and eliminates many unnecessary redundant calculations.

[0048] Considering that there is no significant periodic feature in the 10HZ data, compared with the original algorithm, the embodiment of the present invention only retains the multi-scale convolutional layer (MIC) in the model, extracts the high-frequency details of the electrical appliance at 10 Hz through the Local module in the MIC, aggregates the time steps through the Global module, captures the long-term operation mode of the electrical appliance, and identifies the unique operation characteristics of different electrical appliances by mixing high-frequency and low-frequency information, thereby achieving decoupling of the operating status of the electrical appliances.

[0049] The specific implementation method is to first manually construct the target electrical appliance superposition use case, then use the superimposed use case as the model input, use the model output and the target electrical appliance waveform before superposition to calculate the model loss, and train the neural network to build the MICN model. The high-frequency details of the electrical appliance are extracted, and the time steps are aggregated through the Global module of the MICN model to capture the long-term operation mode of the electrical appliance. By mixing high-frequency and low-frequency information, the unique operation characteristics of different electrical appliances are identified, and the waveforms of the target electrical appliances running at the same time are decoupled to obtain a pure single-type target electrical appliance waveform. Finally, M clusters of target electrical appliances C are obtained. M , the target electrical appliance C M The information includes the first opening time Last closing time Active power sequence Reactive power sequence

[0050] Step S3: feature extraction step, constructing a feature extractor according to the historically accumulated load data features of each target electrical appliance, and using the feature extractor to obtain the feature representation of a single target electrical appliance.

[0051] In the feature extraction step, a feature extractor is constructed according to the historically accumulated load data features of each target electrical appliance. This step specifically includes the following steps: obtaining waveform data of a certain frequency of each type of target electrical appliance as a positive sample, and obtaining waveform data of a certain frequency of target electrical appliances other than this type as a negative sample; inputting the positive sample and the negative sample into a BI-LSTM autoencoder, compressing them into a low-dimensional representation and reconstructing them, using the low-dimensional output of the encoder as the feature of the sample, calculating the contrast loss between the positive sample pair and the negative sample pair, optimizing the sample reconstruction loss and the loss between the positive and negative samples to optimize the BI-LSTM encoder parameters, so that the encoder can become a feature extractor.

[0052] Furthermore, the loss function formula of the BI-LSTM encoder is as follows:

[0053] X={X + ,X -}

[0054] X′={X′ + ,X′ -}

[0055] D′={D′ + ,D′ -}

[0056] Loss = mse(X,X′)+sim(D′) i,+ ,D′ j,+ )-sim(D′ i,+ ,D′ k,- )

[0057] Among them, X is the model input, X + represents a positive sample, X - represents a negative sample, X′ is the output of the autoencoder, D′ is the feature vector after the encoder part is compressed, mse represents the mean square error, and sim represents the cosine similarity.

[0058] Step S4: target appliance classification step, taking the single target appliance feature representation obtained in the feature extraction step and the static input as input, and realizing the prediction of the secondary category of the target appliance through the TFT-C model. In this embodiment, the static input includes the season when the target appliance is running, the temperature when the target appliance is running, and the time when the target appliance starts running.

[0059] like Figure 4 As shown in Figure 1, the TFT-C model is a deep learning model designed for time series forecasting. It combines multiple mechanisms of neural networks to handle complex relationships in time series data. It is used to handle uncertainty and multi-scale dependencies in time series.

[0060] The architecture of the TFT model includes an input layer and an embedding layer. The input layer is used to process different types of inputs, including time series inputs and static inputs, that is, features that do not change over time. The embedding layer embeds and maps the categorical features to transform them into continuous feature representations that can be used by the model. The architecture of the TFT model includes a variable selection network, which is used to dynamically select the most relevant input features. Time series data often contains a large number of features. TFT dynamically selects the most important features for each time step through the variable selection network. The specific implementation processes each input feature separately through a gated residual network to calculate the importance weight of the feature. The architecture of the TFT model includes an LSTM encoder / decoder to learn the sequential information and long-term dependencies of time series data. The specific implementation uses a bidirectional long short-term memory network for encoding to enhance feature expression by capturing the previous and next information; the decoder uses a unidirectional LSTM to predict future time steps. The architecture of the TFT model includes a self-attention mechanism to capture long-term dependencies and global relationships in time series. The specific implementation introduces a multi-head self-attention mechanism to enable the model to focus on the relationships and patterns between different time steps, rather than just local time dependencies. Gated residual network is used to learn complex feature relationships through residual connections, while using gating mechanisms to control the flow of information. GRN contains fully connected layers, nonlinear activation functions, gating mechanisms, and layer normalization, which can learn deeper feature patterns. TFT also contains an explanatory module that can output the importance weight of each feature to explain the model's prediction decisions. By integrating variable selection weights and self-attention weights, it provides a temporal dependency explanation of features and the importance of static features.

[0061] TFT dynamically selects the most important features for each time step, which makes the model more robust when dealing with high-dimensional input and noisy data. By combining LSTM encoder / decoder and self-attention mechanism, TFT is able to capture dependencies on different time scales. Compared with traditional black box models, TFT provides a certain degree of model interpretability through variable selection network and attention mechanism to help understand the decision-making process of the model. At the same time, TFT can be used to process various types of time series data, including but not limited to multivariate, multi-step forecasts and sequences with missing values.

[0062] The embodiment of the present invention uses the 10HZ active power and reactive power data during operation as time series input, and uses the season when the target appliance is running, the temperature when the target appliance is running, and the time when the target appliance starts running as static input. At the model level, compared with the original TFT model, the present invention replaces the self-attention layer in TFT with a patch self-attention layer. By dividing the 10HZ time series data into patches, while enhancing the semantic information of the time series data, the computational complexity of the self-attention layer is greatly reduced. Finally, an MLP and softmax are used to complete the secondary classification of the target appliance. Assume that the active power and reactive power sequences are Static input includes the season when the target appliance is running, the temperature when the target appliance is running, and the time when the target appliance starts running. The calculation formula of the TFT-C model is as follows:

[0063]

[0064] ”'

[0065] T″=Patch(T′)

[0066] O=Self-attention(T″,S)

[0067]

[0068] Output = Softmax(O)

[0069] in, is the time when the target appliance is turned on, Variable_select() is the variable selection network, is the active power sequence of the target electrical appliance, is the reactive power sequence of the target appliance, Lstm_encoder() is the LSTM encoder, GRN() is the gated residual network, Patch() and Self-attention() are attention mechanisms used to divide the time series data of a certain frequency, MLP() is a multi-layer perceptron, Softmax() is a normalized exponential function, and Output represents the secondary category of the target appliance that is finally output.

[0070] Output can analyze the weights in the Variable_select() module to obtain the importance of static inputs such as temperature, time, and season, so as to adjust the type of static input in a targeted manner during the subsequent model training process. The attention weights in self_attention can be visualized to obtain the importance of each operating stage during the operation of the target appliance, providing guidance for the subsequent feature extraction of the target appliance data.

[0071] Step S5: Encoder update step, taking the prediction of the target appliance secondary category output by the target appliance classification step as input, adjusting the parameters of the BI-LSTM encoder, and only considering the reconstruction loss of the sample, extracting the subsequent target appliance features to obtain the personalized feature extraction scheme for the household. The loss function formula is as follows:

[0072] Loss = mse(X,X′)

[0073] Where mse represents mean square error, X is the model input, and X′ is the autoencoder output.

[0074] Step S6: Reclassification step, using the BI-LSTM encoder adjusted in the encoder update step, rematching the target appliances that were not successfully classified in the target appliance classification step to determine the target appliance category to which they belong. Fine-tuning the parameters of the feature extractor through the user's target appliance data realizes adaptive updating of the algorithm parameters, and can build a feature extraction scheme that fits the user's habits to adapt to the differences in electricity usage habits of different household users and complex and diverse appliance usage scenarios, so as to improve the accuracy and adaptability of the secondary classification of the target appliance.

[0075] The advantage of the embodiment of the present invention is that it provides an electrical appliance classification method, which uses the MICN network to achieve decoupling of complex scenes superimposed on target electrical appliances, achieves accurate identification of different types of target electrical appliances, and greatly simplifies the difficulty of subsequent classification of different types of target electrical appliances. The present invention can accurately identify and classify target electrical appliances by combining training feature extractors, with higher accuracy than traditional methods. The present invention can also fine-tune the parameters of the feature extractor through the user's target electrical appliance data, achieve adaptive updating of algorithm parameters, and construct a feature extraction scheme that fits user habits to adapt to the differences in electricity usage habits of different household users and complex and diverse electrical appliance usage scenarios, so as to improve the accuracy and adaptability of the secondary classification of target electrical appliances.

[0076] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for classifying electrical appliances, characterized in that: The specific steps include: The data collection step collects a voltage sequence U and a current sequence I of a total power supply at a certain frequency, and obtains an active power sequence P and a reactive power sequence Q of the total power supply; The target electrical appliance identification step is to obtain information of each target electrical appliance under the total power supply through the active power sequence P and the reactive power sequence Q of the total power supply, and use the MICN model to process the information of each target electrical appliance, decouple and split the coupling relationship of different electrical appliances, so as to obtain the characteristics of a single target electrical appliance; A feature extraction step, constructing a feature extractor according to the historically accumulated load data features of each target electrical appliance, and using the feature extractor to obtain a feature representation of a single target electrical appliance; In the target appliance classification step, the single target appliance feature representation obtained in the feature extraction step is used together with the static input as input, and the secondary category of the target appliance is predicted through the TFT-C model.

2. The electrical appliance classification method according to claim 1, characterized in that: The target electrical appliance identification step specifically includes the following steps: The target electrical appliance event extraction step is to obtain each target electrical appliance h under the total power supply according to the active power sequence P and the reactive power sequence Q. id information, the target appliance h id The information includes the opening time Closing time Active power sequence during operation And the reactive power sequence during operation The first target appliance clustering step is to cluster the target appliances h at a certain time interval. id Filter out the target appliances that are naturally turned off; cluster the target appliances that are naturally turned off so that the time interval between every two adjacent target appliances is greater than a preset time, so as to obtain K groups of target appliances The second target appliance clustering step uses the MICN model to cluster each group of target appliances. Further clustering is performed to finally obtain M clusters of target appliances C M , the target electrical appliance C M The information includes the first opening time Last closing time Active power sequence Reactive power sequence 3. The electrical appliance classification method according to claim 2, characterized in that: In the second target electrical appliance clustering step, the MICN model is used to cluster each group of target electrical appliances. The step of further clustering specifically includes the following steps: Artificially construct target electrical appliance superposition cases, use the superimposed cases as model input, use the model output and the target electrical appliance waveform before superposition to calculate the model loss, train the neural network, and thus construct the MICN model; The high-frequency details of the electrical appliances are extracted, and the time steps are aggregated through the Global module of the MICN model to capture the long-term operation mode of the electrical appliances. By mixing high-frequency and low-frequency information, the unique operation characteristics of different electrical appliances are identified, and the waveforms of the target electrical appliances running simultaneously are decoupled to obtain a pure waveform of a single type of target electrical appliance.

4. The electrical appliance classification method according to claim 1, characterized in that: In the feature extraction step, the step of constructing a feature extractor according to the historically accumulated load data features of each target electrical appliance specifically includes the following steps: The waveform data of a certain frequency of each type of target electrical appliance is obtained as a positive sample, and the waveform data of a certain frequency of target electrical appliances of other types than this type is obtained as a negative sample; the positive and negative samples are input into a BI-LSTM autoencoder, compressed into a low-dimensional representation and reconstructed, the low-dimensional output of the encoder is used as the feature of the sample, the contrast loss between the positive sample pair and the negative sample pair is calculated, and the sample reconstruction loss and the loss between the positive and negative samples are optimized to optimize the BI-LSTM encoder parameters, so that the encoder can become a feature extractor.

5. The electrical appliance classification method according to claim 4, characterized in that: The loss function formula of the BI-LSTM encoder is as follows: X={X + ,X - } X′={X′ + ,X′ - } D′={D′ + ,D′ - } Loss=mse(X,X′)+sim(D′ i,+ ,D′ j,+ )-sim(D′ i,+ ,D′ k,- ) Among them, X is the model input, X + represents a positive sample, X - represents a negative sample, X′ is the output of the autoencoder, D′ is the feature vector after partial compression of the encoder, mse represents mean square error, and sim represents cosine similarity.

6. The electrical appliance classification method according to claim 1, characterized in that: In the target electrical appliance classification step, the calculation formula of the TFT-C model is as follows: T″=Patch(T′) O=Self-attention(T″,S) Output = Softmax(O′) in, is the time when the target appliance is turned on, Variable_select() is the variable selection network, is the active power sequence of the target electrical appliance, is the reactive power sequence of the target appliance, Lstm_encoder() is the LSTM encoder, GRN() is the gated residual network, Patch() and Self-attention() are attention mechanisms used to divide the time series data of a certain frequency, MLP() is a multi-layer perceptron, Softmax() is a normalized exponential function, and Output represents the secondary category of the target appliance that is finally output.

7. The electrical appliance classification method according to claim 1, characterized in that: In the target electrical appliance classification step, the static input includes the season when the target electrical appliance is operating, the temperature when the target electrical appliance is operating, and the time when the target electrical appliance starts operating.

8. The electrical appliance classification method according to claim 1, characterized in that: After the target electrical appliance classification step, the method further includes: In the encoder updating step, the prediction of the secondary category of the target electrical appliance outputted in the target electrical appliance classification step is used as input, the parameters of the BI-LSTM encoder are adjusted, and only the reconstruction loss of the sample is considered. The loss function formula is as follows: Loss = mse(X,X′) Where mse represents mean square error, X is the model input, and X′ is the autoencoder output.

9. The electrical appliance classification method according to claim 8, characterized in that: After the encoder updating step, the method further comprises: The reclassification step uses the BI-LSTM encoder adjusted in the encoder updating step to rematch the target electrical appliances that were not successfully classified in the target electrical appliance classification step to determine the target electrical appliance category to which they belong.

10. A data processing device, characterized in that: include: A memory for storing executable program codes; as well as A processor is used to read the executable program code to run a computer program corresponding to the executable program code to execute the steps in the electrical appliance classification method according to any one of claims 1 to 9.

Images