A power consumption behavior identification method and device

Abstract

The application discloses a power consumption behavior identification method and device, comprising: collecting user power consumption data, preprocessing the user power consumption data to generate a plurality of sub-data curves; the preprocessing operation comprises data cleaning operation and data segmentation operation; obtaining all power consumption data curves, using a multilayer perception to add labels to each data curve to form a matrix of each data curve; obtaining a characteristic matrix of each sub-data curve according to linear projection, calculating a first correlation matrix of each sub-data curve and a characteristic curve of a typical power consumer according to scaling dot product calculation of the characteristic matrix; and performing normalization processing and feedforward neural network processing on all first correlation matrices to generate a second correlation matrix; performing global average pooling calculation on the second correlation matrix according to a global average pooling device, and calculating a possibility probability distribution of operation of each electric appliance in the user power consumption data according to a decision device.

Description

Technical Field

[0001] This invention relates to the field of electricity consumption identification technology, and in particular to a method and apparatus for identifying electricity consumption behavior. Background Technology

[0002] Non-intrusive load identification is a technology for decomposing user loads. It samples the voltage and current of the user's main incoming line and monitors information such as the type, operating status, and power consumption of each user's electrical equipment. A comprehensive and practical load identification technology, through in-depth mining of electricity consumption data, identification of user electricity consumption behavior, and identification of adjustable loads, is of great significance for load forecasting, demand response, guiding users to improve their electricity consumption patterns, rationally scheduling electricity consumption plans, and improving energy efficiency.

[0003] Existing non-intrusive load identification technologies can be broadly categorized into several types. One type is based on transient and steady-state electrical characteristics, identifying appliance types based on differences in harmonic characteristics, power characteristics, VI trajectory characteristics, and electrical characteristics among different appliances. This type requires high computing power from the equipment. Another type is mathematical optimization-based identification methods. The solution efficiency of these load identification algorithms is negatively correlated with the type and number of loads. The third type is based on active and reactive power, employing intelligent identification algorithms to identify appliances, such as Support Vector Machines (SVM), clustering methods, Artificial Neural Networks (ANN), and deep learning. The load identification performance of this type of algorithm is related to the data used to train the method. However, since all three types of non-intrusive load identification technologies are computationally intensive, and new smart meters use low-cost MCUs, their computing power is not suitable for computationally intensive methods. Summary of the Invention

[0004] This invention provides a method and apparatus for identifying electricity consumption behavior, in order to solve the technical problem that existing non-intrusive load identification methods are computationally intensive and unsuitable for new smart meters.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a method for identifying electricity consumption behavior, comprising:

[0006] Collect user electricity consumption data, perform preprocessing operations on the user electricity consumption data, and generate several sub-data curves; the preprocessing operations include data cleaning operations and data segmentation operations.

[0007] All electricity consumption data curves are acquired, and labels are added to each data curve using a multilayer perceptron to form a matrix of data curves; the all electricity consumption data curves include all sub-data curves of user electricity consumption data and characteristic curves of typical electrical appliances;

[0008] The characteristic matrix of each sub-data curve is obtained by linear projection. The first correlation matrix between each sub-data curve and the characteristic curve of a typical appliance is calculated by scaling dot product based on the characteristic matrix. Then, all first correlation matrices are normalized and processed by a feedforward neural network to generate a second correlation matrix.

[0009] The second correlation matrix is ​​calculated using a global average pooler, and the probability distribution of the operation of each appliance in the user's electricity consumption data is calculated using the decision-maker.

[0010] This invention cleanses and segments collected user electricity consumption data to eliminate spikes, glitches, anomalies, and statistical differences in each data stream, resulting in smoother user electricity consumption data. Simultaneously, a multilayer perceptron is used to embed the user electricity consumption data into variable labels, assigning a label to each complete data curve and independently embedding the entire data curve into the variable labels. This embedded label aggregates the complete global properties of the data curve, better extracting its dynamic features. Since convolutional neural networks are not required in the data embedding process, system resources are reduced, system computation is accelerated, and it is more suitable for use in edge computing terminals such as new smart meters.

[0011] Furthermore, the process of collecting user electricity consumption data and preprocessing the data to generate several sub-data curves is as follows:

[0012] The user's electricity consumption data at the user's home entrance is obtained from the metering unit; the user's electricity consumption data includes current information and voltage information.

[0013] The user electricity consumption data is processed to remove outliers, and the user electricity consumption data is normalized according to the sliding window to generate an electricity consumption data curve.

[0014] The electricity consumption data curve is segmented based on the data packet length and time dimension of the user's electricity consumption data to generate several sub-data curves; the data segmentation includes precise segmentation and redundant segmentation.

[0015] Furthermore, the process of acquiring all electricity consumption data curves and using a multilayer perceptron to add labels to each data curve to form a matrix of data curves specifically involves:

[0016] Acquire all sub-data curves of user electricity consumption data and characteristic curves of typical electrical appliances. Add labels to each sub-data curve and characteristic curve according to the multilayer perceptron, acquire the features of all sub-data curves and characteristic curves, and form a matrix of data curves of each sub-data curve and characteristic curve.

[0017] By embedding each sub-data curve and feature curve into variable labels based on the labels, the global characteristics of user electricity consumption data can be obtained.

[0018] Furthermore, the step of obtaining the characteristic matrix of each sub-data curve based on linear projection, calculating the first correlation matrix between each sub-data curve and the characteristic curve of a typical appliance based on the scaling dot product of the characteristic matrix, and then normalizing and processing all the first correlation matrices using a feedforward neural network to generate a second correlation matrix, specifically:

[0019] The feature matrix of each sub-data curve is obtained by linear projection based on the multi-head attention mechanism. The feature matrix includes a query sub-matrix, a key sub-matrix, and a value sub-matrix.

[0020] Based on the characteristic matrix of each sub-data curve, a scaled dot product is performed to calculate the first correlation matrix between each sub-data curve and the characteristic curve of each typical electrical appliance.

[0021] The first correlation matrix of each sub-data curve is normalized to a normal distribution, and the variable features of each data curve are extracted according to the variables to generate the second correlation matrix of the electricity consumption data curve; the second correlation matrix is ​​a feature mapping matrix; the variable features include the change amplitude, period and waveform.

[0022] Furthermore, the step of performing global average pooling calculation on the second relevance matrix based on the global average pooler, and calculating the probability distribution of the operation of each appliance in the user's electricity consumption data based on the decision-maker, specifically involves:

[0023] The feature mapping of the electricity consumption data curve is obtained based on the second correlation matrix, and the similarity score between the feature mapping of the electricity consumption data curve and the feature curve of each typical appliance is calculated based on the global average pooler.

[0024] The similarity score is converted by the decision-maker to obtain the probability distribution of the operation of each electrical appliance in the user's electricity data.

[0025] In a second aspect, the present invention provides an electricity consumption behavior identification device, comprising: a preprocessing module, a data embedding module, a correlation calculation module, and a probability distribution module;

[0026] The preprocessing module is used to collect user electricity consumption data, perform preprocessing operations on the user electricity consumption data, and generate several sub-data curves; the preprocessing operations include data cleaning operations and data segmentation operations.

[0027] The data embedding module is used to acquire all electricity consumption data curves, and to add labels to each data curve using a multilayer perceptron to form a matrix of data curves; the all electricity consumption data curves include all sub-data curves of user electricity consumption data and characteristic curves of typical electrical appliances;

[0028] The correlation calculation module is used to obtain the characteristic matrix of each sub-data curve according to the linear projection, calculate the first correlation matrix between each sub-data curve and the characteristic curve of a typical appliance by performing scaling dot product based on the characteristic matrix, and perform normalization and feedforward neural network processing on all the first correlation matrices to generate the second correlation matrix.

[0029] The probability distribution module is used to perform global average pooling calculation on the second correlation matrix according to the global average pooler, and to calculate the probability distribution of the operation of each appliance in the user's electricity consumption data according to the decision-maker.

[0030] Furthermore, the preprocessing module is specifically used for:

[0031] The user's electricity consumption data at the user's home entrance is obtained from the metering unit; the user's electricity consumption data includes current information and voltage information.

[0032] The user electricity consumption data is processed to remove outliers, and the user electricity consumption data is normalized according to the sliding window to generate an electricity consumption data curve.

[0033] The electricity consumption data curve is segmented based on the data packet length and time dimension of the user's electricity consumption data to generate several sub-data curves; the data segmentation includes precise segmentation and redundant segmentation.

[0034] Furthermore, the data embedding module is specifically used for:

[0035] Acquire all sub-data curves of user electricity consumption data and characteristic curves of typical electrical appliances. Add labels to each sub-data curve and characteristic curve according to the multilayer perceptron, acquire the features of all sub-data curves and characteristic curves, and form a matrix of data curves of each sub-data curve and characteristic curve.

[0036] By embedding each sub-data curve and feature curve into variable labels based on the labels, the global characteristics of user electricity consumption data can be obtained.

[0037] Furthermore, the relevance calculation module is specifically used for:

[0038] The feature matrix of each sub-data curve is obtained by linear projection based on the multi-head attention mechanism. The feature matrix includes a query sub-matrix, a key sub-matrix, and a value sub-matrix.

[0039] The feature matrix of each sub-data curve is obtained by linear projection based on the multi-head attention mechanism. The feature matrix includes a query sub-matrix, a key sub-matrix, and a value sub-matrix.

[0040] Based on the characteristic matrix of each sub-data curve, a scaled dot product is performed to calculate the first correlation matrix between each sub-data curve and the characteristic curve of each typical electrical appliance.

[0041] The first correlation matrix of each sub-data curve is normalized to a normal distribution, and the variable features of each data curve are extracted according to the variables to generate the second correlation matrix of the electricity consumption data curve; the second correlation matrix is ​​a feature mapping matrix; the variable features include the change amplitude, period and waveform.

[0042] Furthermore, the probability distribution module is specifically used for:

[0043] The feature mapping of the electricity consumption data curve is obtained based on the second correlation matrix, and the similarity score between the feature mapping of the electricity consumption data curve and the feature curve of each typical appliance is calculated based on the global average pooler.

[0044] The similarity score is converted by the decision-maker to obtain the probability distribution of the operation of each electrical appliance in the user's electricity data. Attached Figure Description

[0045] Figure 1 This is a schematic flowchart of an electricity consumption behavior recognition method provided in an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the structure of a novel smart meter provided in an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of a data preprocessing module for the electricity consumption behavior recognition method provided in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the encoder operation process for embedding independent markers into each data curve of the electricity consumption behavior recognition method provided in an embodiment of the present invention.

[0049] Figure 5 This is a schematic diagram of an encoder for the electricity consumption behavior recognition method provided in an embodiment of the present invention;

[0050] Figure 6 This is a schematic diagram of a power consumption behavior recognition device provided in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1

[0053] Please refer to Figure 1 , Figure 1 A flowchart illustrating an electricity consumption behavior identification method provided in an embodiment of the present invention includes steps 101 to 104, as detailed below:

[0054] Step 101: Collect user electricity consumption data, perform preprocessing operations on the user electricity consumption data, and generate several sub-data curves; the preprocessing operations include data cleaning operations and data segmentation operations;

[0055] In this embodiment, the electricity consumption information data packet at the user's entrance is obtained through the metering unit, and data cleaning operations such as anomaly removal and normalization are performed on the electricity consumption data. The processed data is then segmented in the time dimension.

[0056] In this embodiment, the process of collecting user electricity consumption data and preprocessing the data to generate several sub-data curves specifically involves:

[0057] The user's electricity consumption data at the user's home entrance is obtained from the metering unit; the user's electricity consumption data includes current information and voltage information.

[0058] The user electricity consumption data is processed to remove outliers, and the user electricity consumption data is normalized according to the sliding window to generate an electricity consumption data curve.

[0059] The electricity consumption data curve is segmented based on the data packet length and time dimension of the user's electricity consumption data to generate several sub-data curves; the data segmentation includes precise segmentation and redundant segmentation.

[0060] Please refer to Figure 2 , Figure 2 This is a schematic diagram of a novel smart meter provided in an embodiment of the present invention.

[0061] In this embodiment, the novel smart meter acquires instantaneous current information at the target user's inlet through current sensing methods such as Rogowski coils, and acquires instantaneous voltage information at the user's inlet through voltage parallel connection equal voltage sensing methods. After assembling the voltage and current information over a certain time period into a user electricity consumption data packet according to a preset data format, it is sent to the AI ​​processor.

[0062] In this embodiment, after receiving the user's electricity consumption data packet, the AI ​​processor verifies whether the data packet format meets the requirements according to the specified data format. The data format specifies that the current and voltage information at the same time are stored in the data packet in a two-dimensional data format, and the data format usually uses the format type FLOT.

[0063] In this embodiment, data cleaning operations such as anomaly removal and normalization are performed on user electricity consumption data that conforms to a preset format.

[0064] In this embodiment, the user's electricity consumption data stream is normalized according to the sliding window. For each input data curve d = [d1, d2, ..., d...], ... S ] T By performing translation and scaling operations, it is transformed into a standard Gaussian distribution, resulting in d' = [d'1, d'2, ..., d']. S ] T The normalization process specifically involves:

[0065]

[0066] Wherein, the value μ d ,σ d ∈R C-1 μ d σ represents the mean. d Represents variance. It represents the Hadamardi (or Hadama) stack.

[0067] In this embodiment, normalization processing eliminates spikes, glitches, abnormal data, and statistical differences between each power consumption data stream, thereby obtaining a smooth and normalized user power consumption data curve, which is a complete data curve of user power consumption data.

[0068] In this embodiment, after normalization, the electricity consumption data curve of the user's electricity consumption data is segmented according to the data packet length and time dimension of the user's electricity consumption data to obtain several sub-data curves.

[0069] In this embodiment, if the data packet length of the user's electricity consumption data is greater than the preset data packet segmentation length, then fine segmentation is used to segment the user's electricity consumption data. If the data packet length of the user's electricity consumption data is not greater than the preset data packet segmentation length, then the data segmentation type is determined based on the electricity consumption characteristics of the user's electricity consumption data. When the user's electricity consumption data contains many high-frequency, short-term transient electricity consumption characteristics or in complex scenarios with a wide variety of electrical appliances, redundant segmentation is used.

[0070] In this embodiment, segmentation is divided into two types according to requirements: precise segmentation and redundant segmentation. The basis for this is the length of the user's electricity data packets. If the packet length is too long, exceeding the preset packet segmentation length, redundant segmentation will further increase the already large amount of user electricity data. Therefore, when the packet length is too long, precise segmentation is used. For example, if the user's electricity data consists of 1000 sampled voltage and current data points over time, precise segmentation with a segmentation length of 100 will result in 10 sub-electricity information data packets of length 100. Redundant segmentation with a redundancy of 5 will result in 10 sub-electricity information data packets of length 110. It can be seen that while redundant segmentation achieves better feature extraction results when the packet length is large, it places a greater computational burden on the central processing chip.

[0071] In this embodiment, when the data packet length is small, i.e., not greater than the preset data packet segmentation length, the data segmentation method can be determined based on the electricity consumption characteristics of the user's electricity consumption data. When the user's electricity consumption data contains many high-frequency, short-duration transient electricity consumption features, or in complex scenarios with a wide variety of electrical appliances, such as factories with stamping machines, shared rental housing, or large shopping malls, the resulting electricity consumption data packets will contain many high-frequency, short-duration transient electricity consumption features. In such cases, selecting a redundant segmentation method to process the data can achieve better feature extraction results.

[0072] In this embodiment, since the number of user power consumption data packets obtained after data segmentation may be large, before feature encoding the segmented user power consumption data packets, the data packets to be feature encoded can be selected according to the length of the data packets. For example, when the length of the data packet is 1000, all data packets can be selected; when the length of the data packet is 10000, 100 samples can be selected every 900 samples, for a total of 10 samples of 100 samples, to obtain data packets with a total length of 1000 samples. This can better identify data packets of loads with long-term power consumption characteristics when the computing power of the embedded system is limited.

[0073] Please refer to Figure 3 , Figure 3 This is a schematic diagram of a module for data preprocessing in the electricity consumption behavior recognition method provided in an embodiment of the present invention.

[0074] In this embodiment, data preprocessing includes an electricity data acquisition module, a data normalization module, and a data segmentation module. The electricity data acquisition module acquires user electricity data packets from the user's home entrance and then transmits these packets to the data normalization module for data cleaning operations such as outlier removal and normalization. Finally, the data packets are transmitted to the data segmentation module, where the processed data is segmented along a time dimension.

[0075] Step 102: Obtain all electricity consumption data curves, and use a multilayer perceptron to add labels to each data curve to form a matrix of data curves; the all electricity consumption data curves include all sub-data curves of user electricity consumption data and characteristic curves of typical electrical appliances;

[0076] In this embodiment, the process of acquiring all electricity consumption data curves and using a multilayer perceptron to add labels to each data curve to form a matrix of data curves specifically involves:

[0077] Acquire all sub-data curves of user electricity consumption data and characteristic curves of typical electrical appliances. Add labels to each sub-data curve and characteristic curve according to the multilayer perceptron, acquire the features of all sub-data curves and characteristic curves, and form a matrix of data curves of each sub-data curve and characteristic curve.

[0078] By embedding each sub-data curve and feature curve into variable labels based on the labels, the global characteristics of user electricity consumption data can be obtained.

[0079] In this embodiment, preprocessed user electricity consumption data is transmitted to a multilayer perceptron (MLP) for embedding as variable tokens. Each complete data curve is assigned a label, and the entire data curve is independently embedded into the variable tokens. This data embedding allows the embedded tokens to aggregate the complete global properties of the data curve, better representing its dynamic characteristics.

[0080] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the encoder operation process for embedding independent markers into each data curve in the electricity consumption behavior recognition method provided in this embodiment of the invention.

[0081] In this embodiment, each complete data curve is assigned a label, and the entire data curve is independently embedded into variable tokens. This embedded token aggregates the complete global properties of the data curve, better representing its dynamic characteristics and allowing the system to focus more on the entire variable process. Subsequently, encoders from various Transformer architectures are integrated. The multi-head attention mechanism is utilized more efficiently to describe the correlations between multivariate processes and capture their feature dependencies. A shared feedforward network independently processes and encodes each variable data curve, learning its characteristics; whereas traditional Transformers require both attention mechanisms and feedforward networks for encoding. Therefore, by using different embedding and encoding methods, and embedding encoders from various Transformer architectures, the multivariate correlation analysis and processing of electricity consumption data curves are achieved more quickly and comprehensively.

[0082] Step 103: Obtain the characteristic matrix of each sub-data curve according to the linear projection, calculate the first correlation matrix between each sub-data curve and the characteristic curve of a typical appliance by performing scaling dot product based on the characteristic matrix; and perform normalization and feedforward neural network processing on all the first correlation matrices to generate the second correlation matrix.

[0083] In this embodiment, the step of obtaining the characteristic matrix of each sub-data curve based on linear projection, calculating the first correlation matrix between each sub-data curve and the characteristic curve of a typical appliance by performing a scaling dot product on the characteristic matrix, and then normalizing and processing all the first correlation matrices using a feedforward neural network to generate a second correlation matrix, specifically:

[0084] The feature matrix of each sub-data curve is obtained by linear projection based on the multi-head attention mechanism. The feature matrix includes a query sub-matrix, a key sub-matrix, and a value sub-matrix.

[0085] Based on the characteristic matrix of each sub-data curve, a scaled dot product is performed to calculate the first correlation matrix between each sub-data curve and the characteristic curve of each typical electrical appliance.

[0086] The first correlation matrix of each sub-data curve is normalized to a normal distribution, and the variable features of each data curve are extracted according to the variables to generate the second correlation matrix of the electricity consumption data curve; the second correlation matrix is ​​a feature mapping matrix; the variable features include the change amplitude, period and waveform.

[0087] In this embodiment, after data embedding is completed, feature maps of each data curve are extracted according to the data encoder. The data encoder includes a multi-head attention module, a first residual standardization module, a feedforward module, and a second residual standardization module.

[0088] Please refer to Figure 5 , Figure 5 This is a schematic diagram of a data encoder for an electricity consumption behavior recognition method provided in an embodiment of the present invention.

[0089] In this embodiment, the data encoder is first trained using the feature curves of various typical electrical appliances with embedded tags to obtain the feature matrix of each typical appliance, and then the initial weight coefficient matrix of each module in the data encoder is obtained. Subsequently, the multi-head attention module identifies the real-time electricity consumption data curves after data cleaning and tag embedding, and uses linear projection to obtain the characteristic matrix of each sub-data curve. The characteristic matrix includes a set of query submatrices (Q), key submatrices (K), and value submatrices (V), and these submatrices are used to calculate the scaled dot product. Specifically:

[0090]

[0091] Here, A represents the correlation between the measured electricity consumption data curve and the characteristic curve (sample curve) of a typical appliance. When the measured curve is highly correlated with the sample curve, the measured curve is assigned a greater correlation to the corresponding appliance, and thus receives greater weight in the interaction between the calculation formula and the value V in the next round.

[0092] In this embodiment, a first correlation matrix for each sub-data curve is generated based on the correlation between the electricity consumption data curve and the characteristic curve of a typical electrical appliance.

[0093] In this embodiment, after the correlation calculation is completed, the correlation values ​​of the first correlation matrix of the electricity consumption data curve are normalized to a normal distribution according to the first residual standardization module, thereby reducing the differences caused by measurement inconsistencies.

[0094] In this embodiment, the complex features of the variables contained in the data curve are extracted based on the feedforward neural network and variable labeling, and a feature mapping matrix of each data curve is generated. The feature mapping matrix includes the correlation values ​​between the electricity consumption data curve and the curves of each typical appliance, i.e., the second correlation matrix. The complex features include amplitude, periodicity, and spectrum, thereby learning and recognizing the characteristics of the electricity consumption data curve.

[0095] In this embodiment, after feature extraction, the correlation values ​​of each feature curve in the second correlation matrix are normalized again according to the second residual standardization module, so that these values ​​are normally distributed, thereby reducing the differences caused by inconsistencies in calculation.

[0096] Step 104: Perform global average pooling calculation on the second correlation matrix based on the global average pooler, and calculate the probability distribution of the operation of each appliance in the user's electricity consumption data based on the decision-maker.

[0097] In this embodiment, the step of performing global average pooling calculation on the second relevance matrix based on the global average pooler, and calculating the probability distribution of the operation of each appliance in the user's electricity consumption data based on the decision-maker, specifically involves:

[0098] The feature mapping of the electricity consumption data curve is obtained based on the second correlation matrix, and the similarity score between the feature mapping of the electricity consumption data curve and each sample curve is calculated based on the global average pooler.

[0099] The similarity score is converted by the decision-maker to obtain the probability distribution of the operation of each electrical appliance in the user's electricity data.

[0100] In this embodiment, global average pooling is performed on each data curve according to the global average pooler. The vectorized view of the global average pooling can force the encoder's output values ​​into orthogonal subspaces of different input categories. Simultaneously, global average pooling significantly reduces the number of parameters compared to fully connected layers, reducing resource usage; it eliminates the need for parameter optimization, thereby avoiding overfitting and improving the system's generalization ability.

[0101] In this embodiment, the output of the global average pooler is a real vector, which typically represents the similarity or confidence score between the measured curve and the typical operating curves of various types of electrical appliances.

[0102] In this embodiment, the real number vector output by the global average pooler of the decision maker is converted, and the similarity score is converted according to the softmax function to transform it into a normalized probability distribution. The probability distribution is the probability that the measured electricity consumption data curve belongs to the operating conditions of various typical electrical appliances.

[0103] In this embodiment, the effectiveness of the identification system is adjusted by changing the threshold. Appliances with a high probability within the threshold range are considered to be in use. An index table of typical appliances with probability rankings is found, and then this index is mapped to the appliance names corresponding to the list of typical appliances. The appliance names are output in descending order of probability, thereby identifying the user's electricity usage behavior.

[0104] Please refer to Figure 6 , Figure 6 A schematic diagram of a power consumption behavior recognition device provided in an embodiment of the present invention includes: a preprocessing module 601, a data embedding module 602, a correlation calculation module 603, and a probability distribution module 604;

[0105] The preprocessing module 601 is used to collect user electricity consumption data, perform preprocessing operations on the user electricity consumption data, and generate several sub-data curves; the preprocessing operations include data cleaning operations and data segmentation operations.

[0106] The data embedding module 602 is used to acquire all electricity consumption data curves, and to add labels to each data curve using a multilayer perceptron to form a matrix of each data curve; the all electricity consumption data curves include all sub-data curves of user electricity consumption data and characteristic curves of typical electrical appliances;

[0107] The correlation calculation module 603 is used to obtain the characteristic matrix of each sub-data curve according to the linear projection, calculate the first correlation matrix between each sub-data curve and the characteristic curve of a typical appliance by performing scaling dot product based on the characteristic matrix, and perform normalization and feedforward neural network processing on all the first correlation matrices to generate the second correlation matrix.

[0108] The probability distribution module 604 is used to perform global average pooling calculation on the second correlation matrix according to the global average pooler, and to calculate the probability distribution of the operation of each electrical appliance in the user's electricity consumption data according to the decision-maker.

[0109] In this embodiment, the preprocessing module is specifically used for:

[0110] The user's electricity consumption data at the user's home entrance is obtained from the metering unit; the user's electricity consumption data includes current information and voltage information.

[0111] The user electricity consumption data is processed to remove outliers, and the user electricity consumption data is normalized according to the sliding window to generate an electricity consumption data curve.

[0112] The electricity consumption data curve is segmented based on the data packet length and time dimension of the user's electricity consumption data to generate several sub-data curves; the data segmentation includes precise segmentation and redundant segmentation.

[0113] In this embodiment, the data embedding module is specifically used for:

[0114] Acquire all sub-data curves of user electricity consumption data and characteristic curves of typical electrical appliances. Add labels to each sub-data curve and characteristic curve according to the multilayer perceptron, acquire the features of all sub-data curves and characteristic curves, and form a matrix of data curves of each sub-data curve and characteristic curve.

[0115] By embedding each sub-data curve and feature curve into variable labels based on the labels, the global characteristics of user electricity consumption data can be obtained.

[0116] In this embodiment, the relevance calculation module is specifically used for:

[0117] The feature matrix of each sub-data curve is obtained by linear projection based on the multi-head attention mechanism. The feature matrix includes a query sub-matrix, a key sub-matrix, and a value sub-matrix.

[0118] Based on the characteristic matrix of each sub-data curve, a scaled dot product is performed to calculate the first correlation matrix between each sub-data curve and the characteristic curve of each typical electrical appliance.

[0119] The first correlation matrix of each sub-data curve is normalized to a normal distribution, and the variable features of each data curve are extracted according to the variables to generate the second correlation matrix of the electricity consumption data curve; the second correlation matrix is ​​a feature mapping matrix; the variable features include the change amplitude, period and waveform.

[0120] In this embodiment, the probability distribution module is specifically used for:

[0121] The feature mapping of the electricity consumption data curve is obtained based on the second correlation matrix, and the similarity score between the feature mapping of the electricity consumption data curve and the feature curve of each typical appliance is calculated based on the global average pooler.

[0122] The similarity score is converted by the decision-maker to obtain the probability distribution of the operation of each electrical appliance in the user's electricity data.

[0123] In this embodiment, data cleaning and segmentation operations are performed on the collected user electricity consumption data to eliminate spikes, glitches, abnormal data, and statistical differences between each data stream, thereby obtaining smooth user electricity consumption data. Simultaneously, a multilayer perceptron is used to embed the user electricity consumption data as variable labels, assigning a label to each complete data curve and independently embedding the entire data curve into the variable labels. This embedded label aggregates the complete global properties of the data curve, better extracting its dynamic features. Since convolutional neural networks are not required in the data embedding, the system's resource requirements are reduced, system computation is accelerated, and it is more suitable for use in edge computing terminals such as new smart meters.

[0124] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for identifying electricity consumption behavior, characterized in that, include: Collect user electricity consumption data, perform preprocessing operations on the user electricity consumption data, and generate several sub-data curves; the preprocessing operations include data cleaning operations and data segmentation operations. All electricity consumption data curves are acquired, and labels are added to each data curve using a multilayer perceptron to form a matrix of data curves. These electricity consumption data curves include all sub-data curves of user electricity consumption data and characteristic curves of typical appliances. Specifically: all sub-data curves of user electricity consumption data and characteristic curves of typical appliances are acquired; labels are added to each sub-data curve and characteristic curve using a multilayer perceptron; the features of all sub-data curves and characteristic curves are acquired, and a matrix of data curves is formed; each sub-data curve and characteristic curve is embedded into variable markers based on the labels to obtain the global features of user electricity consumption data. The feature matrices of each sub-data curve are obtained through linear projection. A first correlation matrix between each sub-data curve and the feature curves of typical electrical appliances is calculated using a scaled dot product based on these feature matrices. All first correlation matrices are then normalized and processed using a feedforward neural network to generate a second correlation matrix. Specifically, the feature matrices of each sub-data curve are obtained using linear projection based on a multi-head attention mechanism. These feature matrices include query submatrices, key submatrices, and value submatrices. A first correlation matrix between each sub-data curve and the feature curves of typical electrical appliances is calculated using a scaled dot product based on these feature matrices. The first correlation matrices of each sub-data curve are normalized to a normal distribution, and variable features of each data curve are extracted based on the variables to generate a second correlation matrix for the electricity consumption data curves. The second correlation matrix is ​​a feature mapping matrix. The variable characteristics include the magnitude of change, period, and waveform; The second relevance matrix is ​​calculated using a global average pooler, and the probability distribution of each appliance's operation in the user's electricity consumption data is calculated using the decision-maker. Specifically, the feature map of the electricity consumption data curve is obtained based on the second relevance matrix, and the similarity score between the feature map of the electricity consumption data curve and the feature curve of each typical appliance is calculated using the global average pooler. The similarity score is then converted using the decision-maker to obtain the probability distribution of each appliance's operation in the user's electricity consumption data.

2. The electricity consumption behavior identification method as described in claim 1, characterized in that, The process involves collecting user electricity consumption data and preprocessing the data to generate several sub-data curves, specifically: The user's electricity consumption data at the user's home entrance is obtained from the metering unit; the user's electricity consumption data includes current information and voltage information. The user electricity consumption data is processed to remove outliers, and the user electricity consumption data is normalized according to the sliding window to generate an electricity consumption data curve. The electricity consumption data curve is segmented based on the data packet length and time dimension of the user's electricity consumption data to generate several sub-data curves; the data segmentation includes precise segmentation and redundant segmentation.

3. An electricity consumption behavior identification device, characterized in that, include: The module includes a preprocessing module, a data embedding module, a relevance calculation module, and a probability distribution module. The preprocessing module is used to collect user electricity consumption data, perform preprocessing operations on the user electricity consumption data, and generate several sub-data curves; the preprocessing operations include data cleaning operations and data segmentation operations. The data embedding module is used to acquire all electricity consumption data curves, add labels to each data curve using a multilayer perceptron, and form a matrix of data curves. All electricity consumption data curves include all sub-data curves of user electricity consumption data and characteristic curves of typical appliances. Specifically: acquire all sub-data curves of user electricity consumption data and characteristic curves of typical appliances; add labels to each sub-data curve and characteristic curve using a multilayer perceptron; acquire the features of all sub-data curves and characteristic curves; and form a matrix of data curves. Based on the labels, embed each sub-data curve and characteristic curve into variable markers to obtain the global features of user electricity consumption data. The correlation calculation module is used to obtain the characteristic matrix of each sub-data curve based on linear projection, calculate the first correlation matrix between each sub-data curve and the characteristic curve of a typical appliance by performing a scaling dot product based on the characteristic matrix, and normalize and process all the first correlation matrices using a feedforward neural network to generate a second correlation matrix. Specifically, it uses a multi-head attention mechanism to obtain the characteristic matrix of each sub-data curve by linear projection, the characteristic matrix including a query sub-matrix, a key sub-matrix, and a value sub-matrix; it performs a scaling dot product based on the characteristic matrix of each sub-data curve to calculate the first correlation matrix between each sub-data curve and the characteristic curve of each typical appliance; it normalizes the first correlation matrix of each sub-data curve to a normal distribution, and extracts the variable features of each data curve based on the variables to generate the second correlation matrix of the electricity consumption data curve; the second correlation matrix is ​​a feature mapping matrix. The variable characteristics include the magnitude of change, period, and waveform; The probability distribution module is used to perform global average pooling calculation on the second relevance matrix based on the global average pooler, and to calculate the probability distribution of the operation of each appliance in the user's electricity consumption data based on the decision-maker. Specifically, it obtains the feature map of the electricity consumption data curve based on the second relevance matrix, calculates the similarity score between the feature map of the electricity consumption data curve and the feature curve of each typical appliance based on the global average pooler, and converts the similarity score based on the decision-maker to obtain the probability distribution of the operation of each appliance in the user's electricity consumption data.

4. The electricity consumption behavior identification device as described in claim 3, characterized in that, The preprocessing module is specifically used for: The user's electricity consumption data at the user's home entrance is obtained from the metering unit; the user's electricity consumption data includes current information and voltage information. The user electricity consumption data is processed to remove outliers, and the user electricity consumption data is normalized according to the sliding window to generate an electricity consumption data curve. The electricity consumption data curve is segmented based on the data packet length and time dimension of the user's electricity consumption data to generate several sub-data curves; the data segmentation includes precise segmentation and redundant segmentation.

Images