A non-intrusive load monitoring method and related apparatus

Abstract

The present application relates to the technical field of power monitoring, and particularly relates to a non-intrusive load monitoring method and related device, which comprises: obtaining an aggregated power sequence to be decomposed to which a target electrical appliance belongs; inputting the aggregated power sequence into a pre-trained non-intrusive load monitoring model to obtain a power decomposition result of the target electrical appliance. In the model, the classification branch, the shared branch and the regression branch adopt a task enhancement mechanism of first fusion and then weighting, and under the condition of lacking real power labels of single electrical appliances, the non-intrusive load monitoring model introduces multi-source weak supervision labels, physical consistency joint loss and high-confidence pseudo-label self-distillation iteration during training, thereby improving the accuracy, rationality and generalization ability of power decomposition of the target electrical appliance, and overall solving the technical problem of low monitoring accuracy of existing non-intrusive load monitoring methods.

Description

Technical Field

[0001] This invention relates to the field of power monitoring technology, and in particular to a non-intrusive load monitoring method and related device. Background Technology

[0002] Non-intrusive load monitoring (NIMBY) refers to a technology that identifies and decomposes the operating status and power consumption of individual target appliances by utilizing aggregated power, current, or voltage signals obtained from the user's main meter, smart meter, or other single-point data acquisition devices, without installing independent sensors or sub-meters for each target appliance. Compared to invasive monitoring methods, NIMBY does not require large-scale modifications to the existing power distribution system and has advantages such as convenient installation, lower data acquisition costs, and wider applicability. Therefore, it has high application value in fields such as smart homes, building energy management, industrial equipment monitoring, demand response, energy-saving diagnostics, and integrated energy services.

[0003] Existing non-intrusive load monitoring methods can be broadly categorized into three types. The first type is based on manual feature and rule-based reasoning. These methods typically extract features such as step changes, steady-state plateaus, rising edges, falling edges, and periodic fluctuations from the aggregate power curve, and combine these with threshold judgment, template matching, or state machine rules to identify target electrical appliances. While this approach is relatively intuitive and effective in simple power consumption scenarios, its accuracy and stability often significantly decrease when multiple appliances operate simultaneously, load changes are superimposed, background noise is strong, or equipment operating modes are complex.

[0004] The second category is based on probabilistic models and optimization solutions. These methods typically employ hidden Markov models, conditional random fields, combinatorial optimization, sparse representation, and dictionary learning to model the relationship between aggregated power and individual appliance states or individual appliance power. While these methods offer some theoretical interpretation, they generally rely heavily on prior assumptions, are sensitive to parameter settings, and tend to suffer from high computational complexity, low solution efficiency, and insufficient model transferability in long-term, multi-device, and large-scale application scenarios.

[0005] The third category is based on deep learning methods. With the development of deep learning technology, an increasing number of non-intrusive load monitoring solutions employ convolutional neural networks, recurrent neural networks, temporal convolutional networks, attention networks, or combinations thereof to perform end-to-end modeling of aggregated power sequences and directly output the state information or power estimation results of the target electrical appliances. Compared with traditional methods, deep learning-based methods can automatically learn complex temporal feature representations, thus demonstrating better decomposition capabilities in various public datasets and real-world scenarios.

[0006] To further improve load decomposition performance, some existing deep learning methods have introduced multi-task learning, simultaneously constructing state recognition and power regression tasks. The classification branch outputs the state probability of the target appliance, while the regression branch outputs the power estimate. The power output is then modulated using state information, enhancing the model's output rationality. Meanwhile, to reduce reliance on time-by-time accurate power labels for individual appliances during training, existing techniques have proposed using weak or indirect supervision. For example, target appliance switch status labels, operational log information, or total power consistency constraints can replace direct power labels, mitigating the cost burden of invasive data acquisition. However, models obtained through this approach suffer from low load decomposition accuracy, leading to low accuracy in non-invasive load monitoring. Summary of the Invention

[0007] This invention provides a non-invasive load monitoring method and related apparatus to solve the technical problem of low monitoring accuracy in existing non-invasive load monitoring methods.

[0008] This invention provides a non-invasive load monitoring method, comprising:

[0009] Obtain the aggregated power sequence to which the target electrical appliance belongs;

[0010] The aggregated power sequence is input into a pre-trained non-intrusive load monitoring model to obtain the power decomposition results of the target electrical appliance;

[0011] The non-intrusive load monitoring model includes a classification branch, a shared branch, a regression branch, and a gated fusion layer. The shared branch extracts shared features from the aggregated power sequence. The classification branch extracts classification features from the aggregated power sequence, fuses the classification features and the shared features, and performs a classification task weighting to obtain the state probability of the target appliance. The regression branch extracts regression features from the aggregated power sequence, fuses the regression features and the shared features, and performs a regression task weighting to obtain the power estimation data of the target appliance. The gated fusion layer generates and outputs the power decomposition result of the target appliance based on the state probability and the power estimation data.

[0012] The non-intrusive load monitoring model is trained using the multi-source weak supervision label set of the target electrical appliance and the weak supervision joint loss function, based on a high-confidence pseudo-label generation and self-distillation iterative mechanism.

[0013] Optionally, the process of fusing the classification features and the shared features and weighting the classification task to obtain the state probability of the target electrical appliance includes:

[0014] Using a preset first fusion method, the classification features and the shared features are combined to form a classification-side joint representation, resulting in a classification fusion feature;

[0015] The classification fusion features are weighted by acquiring and utilizing classification weights to obtain a weighted classification representation;

[0016] Based on the weighted classification representation, output the state probability of the target electrical appliance.

[0017] Optionally, the process of fusing the regression features and the shared features and weighting the regression task to obtain the power estimation data of the target electrical appliance includes:

[0018] Using a preset second fusion method, the regression features and the shared features are combined to form a regression-side joint representation, resulting in regression fusion features;

[0019] The regression fusion features are weighted by acquiring and utilizing the regression weights to obtain a weighted regression representation;

[0020] Based on the weighted regression representation, output the power estimation data of the target electrical appliance.

[0021] Optionally, generating and outputting the power decomposition result of the target electrical appliance based on the state probability and the power estimation data includes:

[0022] The power decomposition result of the target electrical appliance is obtained by multiplying the state probability and the power estimation data by element-wise multiplication.

[0023] Optionally, the training steps of the non-invasive load monitoring model include:

[0024] Acquire and use historical aggregate power sequences as input samples;

[0025] Acquire and construct a multi-source weak supervision tag set based on the multi-source weak supervision information of the target electrical appliance;

[0026] Construct an initial monitoring model and a joint loss function with weak supervision;

[0027] Using the weakly supervised joint loss function as a constraint, the initial monitoring model is trained using the input samples and the multi-source weakly supervised label set to obtain the target monitoring model;

[0028] A first target sample is obtained, and the first target sample is input into the target monitoring model to obtain the target output result. Then, a pseudo-label sample is generated using the sample data corresponding to the target output result that meets the preset high confidence condition.

[0029] Obtain the original weakly supervised samples, merge the original weakly supervised samples with the pseudo-labeled samples to obtain the second target samples, and use the second target samples to train the target monitoring model to obtain the non-intrusive load monitoring model.

[0030] Optionally, the aggregated power sequence includes multiple power segments with preset sliding window lengths;

[0031] The step of inputting the aggregated power sequence into the non-intrusive load monitoring model to obtain the power decomposition result of the target electrical appliance includes:

[0032] Each of the power segments is sequentially input into the non-intrusive load monitoring model to obtain the power decomposition results of the target electrical appliance.

[0033] After inputting the aggregated power sequence into the non-intrusive load monitoring model to obtain the power decomposition result of the target electrical appliance, the process further includes:

[0034] The power decomposition results are reconstructed to obtain the continuous power decomposition results of the target electrical appliance.

[0035] Optionally, the weakly supervised joint loss function includes at least two of the following: aggregate consistency loss, state loss, over-limit penalty, non-negativity constraint, rated power ratio constraint, and event consistency loss.

[0036] Another aspect of the present invention provides a non-invasive load monitoring device, comprising:

[0037] The acquisition module is used to acquire the aggregated power sequence to which the target electrical appliance belongs;

[0038] The monitoring module is used to input the aggregated power sequence into a pre-trained non-intrusive load monitoring model to obtain the power decomposition results of the target electrical appliance;

[0039] The non-intrusive load monitoring model includes a classification branch, a shared branch, a regression branch, and a gated fusion layer. The shared branch extracts shared features from the aggregated power sequence. The classification branch extracts classification features from the aggregated power sequence, fuses the classification features and the shared features, and performs a classification task weighting to obtain the state probability of the target appliance. The regression branch extracts regression features from the aggregated power sequence, fuses the regression features and the shared features, and performs a regression task weighting to obtain the power estimation data of the target appliance. The gated fusion layer generates and outputs the power decomposition result of the target appliance based on the state probability and the power estimation data.

[0040] The non-intrusive load monitoring model is trained using the multi-source weak supervision label set of the target electrical appliance and the weak supervision joint loss function, based on a high-confidence pseudo-label generation and self-distillation iterative mechanism.

[0041] Another aspect of the present invention provides an electronic device, the device including a processor and a memory;

[0042] The memory is used to store program code and transmit the program code to the processor;

[0043] The processor is used to execute the method described above according to the instructions in the program code.

[0044] In another aspect, the present invention provides a computer-readable storage medium for storing program code for performing the method described above.

[0045] As can be seen from the above technical solutions, the present invention has the following advantages:

[0046] This invention provides a non-intrusive load monitoring method that obtains the aggregated power sequence to which a target appliance belongs; inputs the aggregated power sequence into a non-intrusive load monitoring model including a classification branch, a shared branch, a regression branch, and a gated fusion layer to obtain the power decomposition result of the target appliance. The shared branch extracts shared features from the aggregated power sequence; the classification branch extracts classification features from the aggregated power sequence, fuses the classification features and the shared features, and performs a classification task weighting to obtain the state probability of the target appliance; the regression branch extracts regression features from the aggregated power sequence, fuses the regression features and the shared features, and performs a regression task weighting to obtain the power estimation data of the target appliance; the gated fusion layer generates and outputs the power decomposition result of the target appliance based on the state probability and the power estimation data. This improves the differentiated utilization of shared information by the classification and regression tasks, overcomes the coarse utilization of shared information in existing multi-task structures, and thus improves the load decomposition accuracy of the non-intrusive load monitoring model, thereby enhancing the accuracy of non-intrusive load monitoring. Furthermore, the non-intrusive load monitoring model of this invention uses a multi-source weakly supervised label set and a weakly supervised joint loss function for the target electrical appliance. It is trained based on a high-confidence pseudo-label generation and self-distillation iterative mechanism, which reduces the dependence on invasive acquisition methods and overcomes the shortcomings of existing supervised non-intrusive load monitoring methods, such as high training costs and difficult engineering implementation. At the same time, it improves the model accuracy and generalization ability under weak supervision conditions. Moreover, by using the weakly supervised joint loss function as a constraint on the model, the output results of the model are more consistent with the physical laws of the actual power system, which improves the recognition accuracy and reduces the boundary drift and local fluctuations in complex electrical appliance scenarios.

[0047] Therefore, the non-invasive load monitoring method provided by this invention solves the technical problem of low monitoring accuracy in existing non-invasive load monitoring methods. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating a non-invasive load monitoring method provided in an embodiment of the present invention;

[0050] Figure 2A schematic diagram of the network structure of the non-intrusive load monitoring model provided in this embodiment of the invention;

[0051] Figure 3 A schematic diagram illustrating the training steps of the non-intrusive load monitoring model provided in this embodiment of the invention;

[0052] Figure 4 A flowchart illustrating a non-intrusive load monitoring method provided as an application example of the present invention;

[0053] Figure 5 This is a structural block diagram of a non-invasive load monitoring device provided in an embodiment of the present invention. Detailed Implementation

[0054] Based on existing technologies, non-intrusive load monitoring technologies have evolved from early methods relying on manual feature and rule-based judgments to methods utilizing deep learning for automatic feature extraction and joint modeling. They have also begun to address issues such as state recognition, power estimation, and weakly supervised training in complex scenarios. However, in practical applications involving complex electrical appliances, multiple devices operating simultaneously, and difficulties in tag acquisition, existing technologies still have room for further improvement and refinement.

[0055] Specifically, the existing technology has at least the following shortcomings.

[0056] (1) Traditional supervisory non-intrusive load monitoring methods require real power tags for individual electrical appliances at each time interval. However, such tags often require intrusive data collection, resulting in high data acquisition costs, difficulty in expanding the data scale, and difficulty in supporting large-scale promotion and application.

[0057] (2) Existing weak supervision methods mainly rely on single supervision information such as switch status labels, and fail to systematically integrate multi-source weak supervision information such as operation logs, rated power, statistical energy consumption, and running time, resulting in insufficient utilization of supervision information.

[0058] (3) Although existing multi-task learning schemes introduce classification and regression branches, the shared features are usually used as a common front end for different tasks. There is a lack of a differentiated recalibration mechanism for the fused task features, which makes the use of shared information by classification and regression tasks inflexible.

[0059] (4) Existing methods are not sufficiently constrained in terms of physical rationality, and are prone to problems such as the predicted total power being greater than the actual aggregated power, the predicted power of a single appliance being negative, and the predicted power allocation of the target appliance not conforming to the prior of the rated power.

[0060] (5) Most existing weak supervision methods only utilize weak labels in a single training round, lacking a high-confidence pseudo-label generation and self-distillation iteration mechanism, and thus cannot further explore the value of the model's high-confidence output.

[0061] To address the aforementioned issues, this invention presents a non-intrusive load monitoring method based on a three-branch fusion and weighting approach with weak supervision constraints. In the absence of real power labels for individual appliances at each time step, it uses aggregated power time series as input and information such as switch status, operation logs, rated power, statistical energy consumption, and runtime as weak supervision constraints. By constructing three parallel paths—a classification branch, a shared branch, and a regression branch—it models the target appliance's state identification features, shared time series features, and power regression features separately. A task enhancement mechanism of first fusion and then weighting is employed to generate enhanced representations for classification and regression tasks, respectively. The state probability and power estimate are then output, and the final power output of the target appliance is obtained through gated fusion. Furthermore, multi-source weak supervision labels, physical consistency joint loss, and high-confidence pseudo-label self-distillation iteration are introduced. The model is optimized by combining the weak supervision joint loss function and the high-confidence pseudo-label self-distillation mechanism, thereby improving the accuracy, rationality, and generalization ability of the target appliance's power decomposition and achieving non-intrusive load monitoring in complex appliance scenarios.

[0062] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0063] Please see Figure 1 The present invention provides a non-invasive load monitoring method, comprising:

[0064] 101. Obtain the aggregate power sequence to be decomposed of the target electrical appliance.

[0065] It should be noted that target appliances refer to appliances that require load monitoring, and there can be one or more of them, such as refrigerators, washing machines, dishwashers, and other appliances with long-cycle, multi-stage, and complex operating states. The aggregated power sequence to be decomposed to which the target appliances belong refers to a power sequence that includes the power information of the target appliances. It can be understood that the aggregated power sequence can be power data obtained under the operation of multiple target appliances.

[0066] In one example, the aggregated power sequence can be power data obtained from data acquisition and preprocessing at the user's main meter, smart meter, home energy management terminal, or other single-point acquisition location.

[0067] In one embodiment, the aggregated power sequence includes multiple power segments with preset sliding window lengths. The step of obtaining the aggregated power sequence may include:

[0068] First, aggregated power time series data is collected at the target collection point. The target collection point can be a user's main meter, a smart meter, a home energy management terminal, or other single-point data collection location of other appliances.

[0069] Secondly, the polymerization power time series is preprocessed to obtain the polymerization power sequence.

[0070] It should be noted that, because historical aggregated time series are preprocessed during the training phase of the non-intrusive load monitoring model to improve the quality of input signals and the stability of model training, the same preprocessing operations are performed on the real-time aggregated power time series when using the trained non-intrusive load monitoring model. These preprocessing operations include one or more of the following: missing value imputation, outlier correction, resampling, normalization, and sliding window segmentation.

[0071] Specifically, the aggregated power time series is divided into multiple initial power segments according to a sliding window of length W and sliding step size S. Then, each initial power segment is normalized to obtain multiple power segments, which are used as input data for the model in step 102.

[0072] It is understood that the processing of input data during the training phase of the non-intrusive load monitoring model is consistent with the processing of input data actually used after training. The following will use the processing of input data during the training phase as an example to explain the workflow of the non-intrusive load monitoring model of the present invention.

[0073] Specifically, let the aggregated power time series used for training be:

[0074] (1)

[0075] in, This represents the aggregated power value at the T-th sampling time, where T represents the total sequence length.

[0076] Among them, the aggregated power time series reflects the total load change after multiple electrical appliances are superimposed, and is the original input for the non-intrusive load decomposition of this invention.

[0077] In one example, a sliding window of length W and a sliding step size S is used to divide the aggregate power time series into input samples. The nth input sample is represented as:

[0078] (2)

[0079] Where W is the length of the input sample window, and n is the index of the window's starting position.

[0080] To reduce the impact of different users, different time periods, and different units of measurement on model training and use, the power segments obtained by sliding window segmentation are standardized to obtain normalized data, which is used as the input data for the model in step 102.

[0081] The normalized aggregate power time series (or the input samples used for training) is included.

[0082] (3)

[0083] in, For the nth input sample after normalization, This represents the mean of the training set samples. Indicates the standard deviation of the training set samples. To prevent extremely small constants with a denominator of zero.

[0084] 102. Input the aggregated power sequence into a pre-trained non-intrusive load monitoring model to obtain the power decomposition results of the target electrical appliance. The non-intrusive load monitoring model includes a classification branch, a shared branch, a regression branch, and a gated fusion layer. The shared branch is used to extract shared features from the aggregated power sequence. The classification branch is used to extract classification features from the aggregated power sequence, fuse the classification features and shared features, and weight the classification task to obtain the state probability of the target electrical appliance. The regression branch is used to extract regression features from the aggregated power sequence, fuse the regression features and shared features, and weight the regression task to obtain the power estimation data of the target electrical appliance. The gated fusion layer is used to generate and output the power decomposition results of the target electrical appliance based on the state probability and power estimation data. The non-intrusive load monitoring model is trained using a multi-source weakly supervised label set of the target electrical appliance, a weakly supervised joint loss function, and a high-confidence pseudo-label generation and self-distillation iterative mechanism.

[0085] In one embodiment, fusing classification features and shared features and weighting the classification task to obtain the state probability of the target appliance includes: using a preset first fusion method to form a classification-side joint representation of classification features and shared features to obtain classification fusion features; obtaining and using classification weights to weight the classification fusion features to obtain a weighted classification representation; and outputting the state probability of the target appliance based on the weighted classification representation.

[0086] In one embodiment, fusing regression features and shared features and weighting regression tasks to obtain power estimation data for the target appliance includes: using a preset second fusion method to form a joint regression-side representation of regression features and shared features to obtain regression fusion features; acquiring and using regression weights to weight the regression fusion features to obtain a weighted regression representation; and outputting power estimation data for the target appliance based on the weighted regression representation.

[0087] In one embodiment, generating and outputting the power decomposition results of the target appliance based on state probabilities and power estimation data includes:

[0088] The power decomposition result of the target electrical appliance is obtained by multiplying the state probability and power estimation data by element-wise multiplication.

[0089] It should be noted that the specific network structure of the non-intrusive load monitoring model in this embodiment is as follows: Figure 2 As shown, it includes classification branches, shared branches, regression branches, and a gated fusion layer. It takes the aggregated power sequence as input and the power decomposition result of the target appliance as output. Therefore, after obtaining the aggregated power sequence, the corresponding power decomposition result can be obtained by inputting the aggregated power sequence into the non-intrusive load monitoring model.

[0090] It is understandable that the internal workflow of the non-intrusive load monitoring model is the same in both the training and actual use phases. Therefore, this embodiment will explain the workflow of the non-intrusive load monitoring model based on the aggregated power time series represented by formula (1) in step 101.

[0091] S1. After obtaining the normalized aggregated power sequence, the aggregated power sequence is simultaneously input into the classification branch, the shared branch, and the regression branch. The classification branch, shared branch, and regression branch have different tasks. The classification branch extracts features related to the state discrimination of the target electrical appliance, focusing on information such as start-stop boundaries, state transitions, operating stages, and active segments. The shared branch extracts shared temporal features that both the classification and regression tasks depend on, focusing on local plateau segments, periodic segments, waveform patterns, and multi-stage transition features. The regression branch extracts features related to the power estimation of the target electrical appliance, focusing on power amplitude changes, waveform trends, and temporal continuity. In this embodiment, the classification branch, shared branch, and regression branch output classification features based on the received aggregated power sequence. Shared features and regression features Among them, classification features mainly characterize the discrimination information of the target electrical appliance in the state recognition task, shared features mainly characterize the basic time series information that both the classification task and the power regression task depend on, and regression features mainly characterize the numerical change information of the target electrical appliance in the power estimation task.

[0092] Let the feature extraction mappings of the classification branch, the shared branch, and the regression branch be respectively... , and Normalized input samples If we input the classification branch, the shared branch, and the regression branch simultaneously, their outputs will be as follows:

[0093] (5)

[0094] in, , and These represent the feature extraction mappings for the classification branch, the shared branch, and the regression branch, respectively.

[0095] In a preferred embodiment, the classification branch, the sharing branch, and the regression branch can all be implemented using a one-dimensional convolutional neural network, such as... Figure 2 As shown, it includes convolutional layers 1 to 5 and fully connected layers 1 to 2. For any branch... Where c represents the classification branch, s represents the shared branch, and r represents the regression branch, the output of the m-th channel of the l-th layer of this branch at time t can be expressed as:

[0096] (6)

[0097] in, This represents the convolution operation. Represents the convolution kernel parameters. Indicates the bias term. This represents the activation function.

[0098] In another preferred embodiment, the classification branch, the sharing branch, and the regression branch may also employ temporal convolutional networks, recurrent neural networks, Transformer networks, or combinations thereof.

[0099] S2. Construct classification-side fusion representations and regression-side fusion representations respectively. That is, using the preset first fusion method, the classification features output by the classification branch are... Shared characteristics with shared branch output By performing fusion, classification fusion features are obtained. Simultaneously, the preset second fusion method will incorporate the regression features output by the regression branch. Shared characteristics with shared branch output By performing fusion, regression fusion features are obtained. Correspondingly, the classification fusion features and regression fusion features are expressed as follows:

[0100] (7)

[0101] in, Represents the classification fusion function. This represents the regression fusion function.

[0102] In a preferred embodiment, the classification fusion function and the regression fusion function are implemented using a feature concatenation and mapping layer approach, namely:

[0103] (8)

[0104] in, This indicates a feature concatenation operation. and These represent the mapping matrices for the classification and regression sides, respectively. and These represent the bias terms, This represents the activation function.

[0105] In other embodiments, the classification fusion function and the regression fusion function can also employ weighted summation, gated fusion, attention fusion, or residual fusion. Any implementation that achieves the technical concept of forming a joint representation on the classification side from state classification features and shared temporal features, and a joint representation on the regression side from power estimation features and shared temporal features, falls under the category of equivalent implementations of this invention.

[0106] S3. After completing the classification and fusion features and regression fusion features Following the construction, this invention further applies task-related weighting to the two types of fused features. Let the classification weight parameters and regression weight parameters be respectively... and The classification weights are then obtained through the Sigmoid mapping. and regression weights :

[0107] (9)

[0108] in, This is the Sigmoid function, used to constrain weight values ​​to the range of 0 to 1.

[0109] Subsequently, the classification fusion features and regression fusion features were weighted separately to obtain the weighted classification representation and the weighted regression representation:

[0110] (10)

[0111] in, This represents element-wise multiplication. For weighted classification fusion features, The weighted regression fusion features.

[0112] As can be seen from the above processing, this embodiment does not simply allocate the shared features themselves, but enhances the fused task representations separately, so that the classification task and the regression task can utilize the shared information more effectively according to their respective objectives.

[0113] Through the aforementioned feature fusion and task-related weighting steps, this invention forms a technical path that differs from existing common shared branch structures. Existing technologies typically send shared information directly to different tasks at the shared feature level, or simply output state and power separately through a simple branch structure. In contrast, this invention, through a method of "first fusing task-specific features with shared features, and then independently recalibrating the fused representation," enables classification and regression tasks to obtain enhanced representations more suitable for their respective objectives. This not only improves the utilization efficiency of shared information but also enhances the synergistic effect between state recognition and power estimation in complex electrical appliance scenarios.

[0114] S4. After completing feature fusion and task-related weighting, output the state probability of the target electrical appliance based on weighted classification, and output the power estimate of the target electrical appliance based on weighted regression.

[0115] The status output can be represented as:

[0116] (11)

[0117] In the formula, This represents the state probability output of the target electrical appliance. The weight parameters for the status output header. For bias terms, This is the Sigmoid function. State probability characterizes the likelihood that a target appliance is in an active state at a given time. A higher probability value indicates that the target appliance is more likely to be in an active state; a lower probability value indicates that the target appliance is more likely to be in an active state.

[0118] Power output can be expressed as:

[0119] (12)

[0120] In the formula, This represents the estimated power of the target electrical appliance. Weighting parameters of the power output head This is the bias term. Power output is used to characterize the power of the target appliance at a given moment.

[0121] S5. To ensure that the final output simultaneously possesses state discrimination constraints and power numerical representation capabilities, this invention further employs a gating fusion mechanism to combine state probabilities with power estimates to obtain the final power output of the target electrical appliance.

[0122] (13)

[0123] in, Indicates the final power output of the target electrical appliance. This indicates element-wise multiplication.

[0124] Through the above-mentioned gating fusion, when the target appliance is in the on state, the state probability is relatively high, and the final power output is mainly determined by the power estimate; when the target appliance is in the off state, the state probability is relatively low, and the final power output is suppressed, thereby reducing the occurrence of false power output when the target appliance is not in operation and improving the physical rationality of the results.

[0125] In one embodiment, a standby power term may be introduced for target electrical appliances that have standby power. The final output can then be represented as:

[0126] (14)

[0127] in, This indicates the standby power of the target appliance. By introducing a standby power term, the target appliance can maintain a more realistic low power output even when it is off.

[0128] Therefore, this step, through the joint design of state output, power output, and gating fusion, enables the model not only to determine whether the target electrical appliance is in operation, but also to output a power value that matches the state, thereby improving the accuracy and rationality of non-intrusive load monitoring results.

[0129] In one embodiment, since the aggregated power sequence can be composed of multiple power segments with preset sliding window lengths, when inputting the aggregated power sequence into a non-intrusive load monitoring model, the power decomposition results of the target appliance can be obtained by sequentially inputting each power segment into the non-intrusive load monitoring model; subsequently,

[0130] The power decomposition results are reconstructed to obtain the continuous power decomposition results of the target electrical appliance.

[0131] It should be noted that in practical use, the aggregated power time series to be decomposed undergoes the same preprocessing and sliding window segmentation as in the training phase, and the processed samples are input into the trained model to obtain the target electrical appliance state probability sequence, power estimation sequence, and final power decomposition results.

[0132] Because there may be overlapping areas between the sliding windows, there is also overlapping data between the power decomposition results. Therefore, for the overlapping areas of the sliding windows, the target electrical power curve under the complete time axis is restored by average reconstruction or weighted reconstruction.

[0133] Let time t be the window set Overlaying reconstructs the output results of multiple windows at that moment to obtain the target electrical appliance power decomposition results under the complete time axis:

[0134] (15)

[0135] in, This represents the estimated power value of the target electrical appliance at time t after reconstruction. Based on this, this embodiment obtains the power decomposition result of the target electrical appliance under the complete time axis.

[0136] This embodiment provides a non-intrusive load monitoring method that obtains the aggregated power sequence to which the target appliance belongs. The aggregated power sequence is input into a non-intrusive load monitoring model comprising a classification branch, a shared branch, a regression branch, and a gated fusion layer to obtain the power decomposition result of the target appliance. The shared branch extracts shared features from the aggregated power sequence; the classification branch extracts classification features from the aggregated power sequence, fuses the classification features and shared features, and weights the classification task to obtain the state probability of the target appliance; the regression branch extracts regression features from the aggregated power sequence, fuses the regression features and shared features, and weights the regression task to obtain the power estimation data of the target appliance; the gated fusion layer generates and outputs the power decomposition result of the target appliance based on the state probability and power estimation data. This improves the differentiated utilization of shared information by the classification and regression tasks, overcomes the coarse utilization of shared information in existing multi-task structures, and thus improves the load decomposition accuracy of the non-intrusive load monitoring model, thereby enhancing the accuracy of non-intrusive load monitoring. Furthermore, the non-intrusive load monitoring model in this embodiment is trained using a multi-source weakly supervised label set of the target electrical appliance and a weakly supervised joint loss function, based on a high-confidence pseudo-label generation and self-distillation iterative mechanism. This reduces the dependence on invasive acquisition methods and overcomes the shortcomings of existing supervised non-intrusive load monitoring methods, such as high training costs and difficult engineering implementation. At the same time, it improves the model accuracy and generalization ability under weak supervision conditions. Moreover, by using the weakly supervised joint loss function as a constraint on the model, the output results of the model are more consistent with the physical laws of the actual power system, improving the identification accuracy and reducing boundary drift and local fluctuations in complex electrical appliance scenarios.

[0137] The above is a description of a non-intrusive load monitoring method provided by an embodiment of the present invention. Based on this embodiment, the training steps of the non-intrusive load monitoring model will be described below.

[0138] In one embodiment, see Figure 3 The training steps for a non-invasive load monitoring model include:

[0139] 201. Obtain and use the historical aggregate power sequence as input samples.

[0140] It should be noted that the historical aggregated power sequence refers to the aggregated power sequence obtained within a preset time period in the past, which is used as the input sample for training the initial monitoring model.

[0141] It is understandable that the historical aggregate power sequence is used as training data and is obtained using the same processing method as in step 101. For details, please refer to step 101, which will not be repeated here.

[0142] 202. Obtain and construct a set of multi-source weak supervision labels based on the multi-source weak supervision information of the target electrical appliance.

[0143] It should be noted that the training phase of this invention does not require the availability of real power labels for individual electrical appliances at every moment. Instead, it constructs a multi-source weakly supervised label set. The multi-source weakly supervised label set includes at least one or more of the following: target appliance switch status label, target appliance start-stop operation log, target appliance rated power information, total energy consumption information within a statistical time interval, target appliance runtime information, and operation plan or process schedule information.

[0144] Let the set of state labels for the i-th target appliance over the entire time series be:

[0145] (16)

[0146] Where T represents the total length of the state sequence, t represents any sampling time, and 1 ≤ t ≤ T. This represents the state information of the i-th target electrical appliance at time t. This state information refers to the switch status label or weak status label of the target electrical appliance at the corresponding sampling time. It should be noted that the status label or weak status label is a type of state-based supervisory information in the multi-source weak supervision label set. It is mainly used to characterize whether the target electrical appliance is in the on state at the corresponding time, and is not equivalent to the entire multi-source weak supervision label set. The multi-source weak supervision label set may also include one or more of the following: target electrical appliance start-stop operation log, rated power information, total energy consumption information within the statistical time interval, target electrical appliance running time information, and operation plan or process schedule information.

[0147] In one embodiment, if manual records or equipment operation logs exist, status labels can be directly constructed based on start-up and shutdown times. In another embodiment, weak status labels can be automatically constructed based on rated power, empirical thresholds, or a small number of known operating segments. In yet another embodiment, if the status information itself is uncertain, soft labels can be constructed instead of using only binary labels. Soft labels are a form of weak status labels, and their values ​​can be between 0 and 1, representing the probability that the target electrical appliance is in the on state when the status information is uncertain.

[0148] Let the rated power of the i-th target electrical appliance be... The total energy consumption of the target appliances within the statistical time interval is The runtime is The above quantities can all be used as auxiliary information in subsequent weak supervision constraints.

[0149] 203. Construct the initial monitoring model and the joint loss function of weak supervision.

[0150] It should be noted that the network structure of the initial monitoring model is consistent with that of the non-intrusive load monitoring model, and the internal workflow is also consistent. Therefore, the description of the initial monitoring model can be found in step 102, and will not be repeated here.

[0151] To enable model training in the absence of real power labels for individual appliances at each time step, this invention constructs a weakly supervised joint loss function to impose multifaceted constraints on the model's output. This weakly supervised joint loss function considers not only the consistency between the predicted total power and the actual aggregated power, but also the accuracy of state identification, the physical rationality of power output, and the consistent relationship between state changes and power changes. This allows the model to learn relatively accurate target appliance state and power characteristics even under weak supervision.

[0152] In one embodiment, the weakly supervised joint loss function includes at least two of the following: aggregate consistency loss, state loss, over-limit penalty, non-negativity constraint, rated power ratio constraint, and event consistency loss.

[0153] Specifically, assuming N target electrical appliances are predicted simultaneously, the total predicted power at time t can be expressed as:

[0154] (17)

[0155] in, This represents the state probability of the i-th target appliance at time t. Let represent the power estimate of the i-th target appliance at time t.

[0156] 1) Aggregate consistency loss.

[0157] The aggregation consistency loss is used to constrain the difference between the sum of the predicted power of all target appliances and the actual aggregate power, and it is defined as:

[0158] (18)

[0159] in, This represents the actual aggregate power value at time t.

[0160] The role of aggregation consistency loss is that, although there are no time-by-time true power labels for individual appliances during the training phase, the sum of the predicted results for all target appliances should be as close as possible to the true total meter power. Through this constraint, the model can use the global information of aggregated power to learn the power decomposition results of individual appliances in reverse, thus forming the main optimization basis under weak supervision.

[0161] 2) State loss.

[0162] The state loss is used to constrain the consistency between the state probabilities output by the model and the weakly supervised state labels, and it is defined as follows:

[0163] (19)

[0164] in, This represents the state label of the i-th target electrical appliance at time t.

[0165] The role of state loss is to supervise the state output branch by utilizing existing state labels, start / stop logs, or soft state information, so that the model can learn more accurately when the target appliance is in a working state and when it is in a non-working state, thereby improving the effectiveness of subsequent gating fusion.

[0166] 3) Penalties for exceeding limits.

[0167] To prevent the sum of the predicted power of all target appliances from exceeding the actual aggregate power, this invention introduces an over-limit penalty term, defined as follows:

[0168] (20)

[0169] The purpose of the over-limit penalty term is to penalize the excess power when the predicted total power exceeds the actual total power, and to set the value to zero when the predicted total power does not exceed the actual total power. This constraint can effectively suppress the situation where the model allocates too much total power during training, thereby improving the physical rationality of the prediction results.

[0170] 4) Non-negative constraint terms.

[0171] To prevent the model from outputting physically unreasonable negative power values, this invention introduces a non-negative constraint term, defined as follows:

[0172] (twenty one)

[0173] in, This represents the final power output of the i-th target appliance at time t.

[0174] The nonnegativity constraint term serves to penalize negative values ​​when the predicted power of a target appliance is less than zero, and sets the term to zero when the predicted power is greater than or equal to zero. This constraint makes the model output more consistent with the physical law of nonnegativity of power in actual electrical equipment.

[0175] 5) Rated power ratio constraint.

[0176] To ensure that the predicted power allocation of each target appliance is consistent with its prior rated power, this invention introduces a rated power proportional constraint term, which is defined as follows:

[0177] (twenty two)

[0178] in, This represents the rated power of the i-th target electrical appliance.

[0179] The purpose of the rated power ratio constraint is to ensure that, when multiple target electrical appliances are operating simultaneously, the distribution ratio of different appliances in the predicted total power should match their prior rated power as closely as possible. This constraint can reduce the problem of power imbalance in the model when multiple devices are operating simultaneously.

[0180] 6) Loss of event consistency.

[0181] To enhance the consistency between the state change boundary and the power change boundary, this invention further introduces an event consistency loss. The state change quantity and the power change quantity are defined as follows:

[0182] (twenty three)

[0183] The event consistency loss is defined as:

[0184] (twenty four)

[0185] in, This represents the power change mapping function, used to map the power change amplitude to the event intensity space.

[0186] The purpose of event consistency loss is to ensure that when the state of the target appliance changes significantly, the power output should also change significantly accordingly; conversely, when the state change is not significant, the power output should not experience excessive jumps. This constraint enhances the consistency between the state boundary and the power boundary, improving the accuracy of boundary identification and phase transition identification in complex appliance scenarios.

[0187] In a preferred embodiment, this embodiment utilizes the above six constraints to construct a total loss function as the weakly supervised joint loss function of the present invention.

[0188] The total loss function is:

[0189] (25)

[0190] in, These are the weight coefficients for each loss term, used to adjust the strength of the effect of different constraint terms during model training.

[0191] This joint loss function allows for multi-faceted constraints on the model by simultaneously utilizing aggregate power information, state information, physical prior information, and event boundary information under weak supervision. This enables the model to both learn aggregate power decomposition relationships and output results that conform to the operating rules of the equipment.

[0192] 204. Using the weakly supervised joint loss function as a constraint, and utilizing the input samples and the multi-source weakly supervised label set, train the initial monitoring model to obtain the target monitoring model.

[0193] It should be noted that during the training process, the input samples are first fed into the classification branch, the shared branch, and the regression branch to obtain classification features, shared features, and regression features; then feature fusion, weight allocation, state output, power output, and gating fusion are completed to obtain the state probability, power estimate, and final power output of the target electrical appliance; based on the true aggregated power obtained in step 201... And the weak supervision joint loss function is calculated from the multi-source weak supervision label set constructed in step 202. The multi-source weak supervision label set includes one or more of the following: weak status labels, start-stop operation logs, rated power information, total energy consumption information within a statistical time interval, target appliance runtime information, and operation plan or process schedule information. The actual aggregated power... This refers to the total load power value obtained by the user's main meter, smart meter, home energy management terminal or other single-point acquisition device in step 101 or step 201 after preprocessing. It is not the real power label of a single appliance at each time step. This real aggregated power reflects the total power consumption of multiple appliances superimposed at the same time. It is one of the input data sources of the non-intrusive load monitoring model, and also serves as the basis for the aggregation consistency constraint in the weakly supervised joint loss function.

[0194] Subsequently, the training optimization module optimized the joint loss function. About the model parameter set Calculate the gradient and update the model parameters using gradient descent or its variants. The update formula is expressed as:

[0195] (26)

[0196] in, This is the learning rate. The joint loss function L represents the set of model parameters. The gradient.

[0197] By repeatedly performing forward propagation, loss calculation, and parameter updates, the parameters of the classification branch, sharing branch, regression branch, fusion module, weight allocation module, and output module gradually converge. When the preset number of training rounds is reached, or the joint loss function meets the preset convergence condition, the model training ends, and the target detection model is obtained.

[0198] After training, the target monitoring model is used for target electrical appliance state identification and power decomposition in the inference phase, and can further serve as the initial model for high-confidence pseudo-label generation and self-distillation iteration.

[0199] Among them, the model parameter set for:

[0200] (27)

[0201] in, , and Let represent the parameter sets of the classification branch, the shared branch, and the regression branch, respectively. and These represent the parameter sets for the classification fusion module and the regression fusion module, respectively. and These represent the weight allocation parameters for the classification and regression sides, respectively. and These represent the parameter sets for the status output head and the power output head, respectively.

[0202] As can be seen from the above, this step iteratively optimizes the model parameters based on the joint loss function to obtain a target model that meets the requirements of non-intrusive load monitoring.

[0203] 205. Obtain the first target sample, input the first target sample into the target monitoring model, obtain the target output result, and generate pseudo-label samples using the sample data corresponding to the target output result that meets the preset high confidence condition.

[0204] It should be noted that the first target sample is the input sample obtained using the same data preprocessing and sliding window segmentation method as in step 101 or step 201. The first target sample can include two categories: the first category is the training sample obtained after preprocessing and sliding window segmentation of the historical aggregated power sequence in step 201; the second category is unlabeled samples containing only the aggregated power sequence and without matched state labels or other weakly supervised labels. In other words, the first target sample itself is a power segment sample input into the target monitoring model, not the model output.

[0205] After obtaining the target monitoring model, the first target sample is input into the model to obtain the target output results. The target output results include the state probability of the target appliance, the power estimate, the final power decomposition result, and the predicted aggregated power obtained by summing the power decomposition results of each target appliance. Subsequently, based on the target output results, it is determined whether the first target sample meets the preset high confidence conditions. The high confidence conditions include at least one of the following: the state probability is higher than a preset state threshold, the aggregation consistency residual between the predicted aggregated power and the actual aggregated power is lower than a preset residual threshold, the predicted power satisfies the non-negativity constraint and the rated power constraint, the power change event is consistent with the operation log record, and the deviation between the predicted energy consumption and the weakly supervised energy consumption label within the statistical time interval is lower than a preset energy consumption deviation threshold.

[0206] For a first target sample that meets the preset high-confidence condition, its target output is converted into a pseudo-label. Specifically, the state probability can be compared with a preset state threshold to generate a pseudo-state label; the final power decomposition result output by the target monitoring model can be used as a pseudo-power label; and pseudo-event labels can be generated based on the change in state probability or power at adjacent time points. Pseudo-labels include one or more of pseudo-state labels, pseudo-power labels, and pseudo-event labels. The first target sample that meets the preset high-confidence condition and its corresponding pseudo-label are combined to obtain a pseudo-label sample.

[0207] 206. Obtain the original weakly supervised samples, merge the original weakly supervised samples and pseudo-labeled samples to obtain the second target samples, and use the second target samples to train the target monitoring model to obtain the non-intrusive load monitoring model.

[0208] It should be noted that the original weakly supervised samples refer to the training samples formed after matching the input samples obtained in step 201 with the multi-source weakly supervised label set constructed in step 202. In other words, the original weakly supervised samples include the input samples obtained by preprocessing and sliding window segmentation of the historical aggregated power sequence, as well as one or more weakly supervised information such as the state weak label, operation log, rated power, statistical energy consumption, runtime, or operation plan information corresponding to the input sample.

[0209] This step merges the original weakly supervised samples with the pseudo-label samples obtained in step 205 to form the second target samples, and uses the second target samples to retrain or fine-tune the target monitoring model, thus forming an iterative optimization process of "initial weakly supervised training - high-confidence pseudo-label generation - self-distillation retraining".

[0210] This invention introduces high-confidence pseudo-labels, which can further mine the effective supervision information contained in the initial output of the model, thereby improving the training accuracy and generalization ability under weak supervision conditions.

[0211] In one application example, see Figure 4The non-intrusive load monitoring method provided by this invention can be implemented as follows: Collect aggregated power data; preprocess the aggregated power data; construct weakly supervised information; extract three-branch features; fuse the extracted three-branch features; weight the features according to task relevance and output the state probability and estimated power; perform gated fusion based on the state probability and estimated power to output the target electrical appliance power; construct a weakly supervised joint loss constraint; generate high-confidence pseudo-labels and iterate self-distillation; output the non-intrusive load monitoring results; and end.

[0212] See Figure 5 The present invention provides a non-invasive load monitoring device, comprising:

[0213] The acquisition module 301 is used to acquire the aggregated power sequence to be decomposed to which the target electrical appliance belongs;

[0214] The monitoring module 302 is used to input the aggregated power sequence into a pre-trained non-intrusive load monitoring model to obtain the power decomposition results of the target electrical appliance;

[0215] The non-intrusive load monitoring model includes a classification branch, a shared branch, a regression branch, and a gated fusion layer. The shared branch extracts shared features from the aggregated power sequence. The classification branch extracts classification features from the aggregated power sequence, fuses the classification features and shared features, and weights the classification task to obtain the state probability of the target appliance. The regression branch extracts regression features from the aggregated power sequence, fuses the regression features and shared features, and weights the regression task to obtain the power estimation data of the target appliance. The gated fusion layer generates and outputs the power decomposition results of the target appliance based on the state probability and power estimation data.

[0216] The non-intrusive load monitoring model is trained using a multi-source weakly supervised label set of the target electrical appliance and a weakly supervised joint loss function, based on a high-confidence pseudo-label generation and self-distillation iterative mechanism.

[0217] The present invention also provides an electronic device, the device including a processor and a memory;

[0218] The memory is used to store program code and transfer the program code to the processor;

[0219] The processor is used to execute the methods of any of the above embodiments according to instructions in the program code.

[0220] The present invention also provides a computer-readable storage medium for storing program code for performing the methods of any of the above embodiments.

[0221] In one application example, this invention provides a non-intrusive load monitoring system. The system includes a data acquisition module, a data preprocessing module, a sample construction module, a weakly supervised label construction module, a classification branch module, a shared branch module, a regression branch module, a classification fusion module, a regression fusion module, a classification weight allocation module, a regression weight allocation module, a state output module, a power output module, a gated fusion module, a joint loss training module, and a pseudo-label generation and self-distillation module. These modules work sequentially and collaboratively to complete the entire non-intrusive load monitoring process from aggregated power input to target appliance power output.

[0222] In summary, the non-invasive load monitoring method and related apparatus provided by the embodiments of the present invention have at least the following beneficial effects:

[0223] (1) In the training phase, this invention does not require the availability of real power labels for individual electrical appliances at each time step. Instead, it uses multi-source weak supervision information such as switch status, operation log, rated power, statistical energy consumption and running time for training, thereby effectively reducing the cost of training data acquisition, reducing the dependence on intrusive acquisition methods, and overcoming the problems of high training cost and difficult engineering implementation of existing supervised non-intrusive load monitoring methods.

[0224] (2) This invention constructs a classification branch, a sharing branch and a regression branch, extracts state discrimination features, shared temporal features and power regression features respectively, and adopts a task enhancement mechanism of first fusion and then weighting to recalibrate the classification fusion representation and the regression fusion representation respectively, thereby improving the ability of the classification task and the regression task to make differentiated use of shared information, and overcoming the problem of rough use of shared information in the existing multi-task structure.

[0225] (3) This invention uses aggregated consistency loss, over-limit penalty, non-negative constraint and rated power ratio constraint to jointly model, so that the prediction results are more in line with the physical laws of the actual power system, thereby overcoming the problems of over-limit prediction of total power, negative prediction power of single electrical appliance and unreasonable power allocation in the existing methods.

[0226] (4) This invention uses event consistency loss to jointly constrain the state change boundary and the power change boundary, thereby improving the recognition accuracy of the target electrical appliance start-stop boundary and the switching position of the operation stage, and alleviating the problems of boundary drift and excessive local fluctuation in complex state electrical appliance scenarios.

[0227] (5) This invention uses a high-confidence pseudo-label generation and self-distillation iteration mechanism to reuse the high-confidence output generated after the initial training of the model for supervised optimization, forming a closed-loop enhancement process, thereby further improving the model accuracy and generalization ability under weak supervision conditions and overcoming the problem of insufficient utilization of single-round training information in existing weak supervision methods.

[0228] (6) This invention is particularly suitable for non-intrusive load monitoring of appliances with long cycles, multiple stages and complex states, such as refrigerators, washing machines and dishwashers. It can improve the accuracy of power decomposition, the rationality of output and engineering applicability while ensuring low tag cost.

[0229] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0230] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0231] Furthermore, in the various embodiments of the present invention, the functional units can be integrated into one processing unit, or each functional unit can be a separate physical entity, or two or more functional units can be integrated into one processing unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0232] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0233] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0234] It should also be noted that in the description of this invention, the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0235] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-invasive load monitoring method, characterized in that, include: Obtain the aggregated power sequence to which the target electrical appliance belongs; The aggregated power sequence is input into a pre-trained non-intrusive load monitoring model to obtain the power decomposition results of the target electrical appliance; The non-intrusive load monitoring model includes a classification branch, a shared branch, a regression branch, and a gated fusion layer. The shared branch extracts shared features from the aggregated power sequence. The classification branch extracts classification features from the aggregated power sequence, fuses the classification features and the shared features, and performs a classification task weighting to obtain the state probability of the target appliance. The regression branch extracts regression features from the aggregated power sequence, fuses the regression features and the shared features, and performs a regression task weighting to obtain the power estimation data of the target appliance. The gated fusion layer generates and outputs the power decomposition result of the target appliance based on the state probability and the power estimation data. The non-intrusive load monitoring model is trained using the multi-source weak supervision label set of the target electrical appliance and the weak supervision joint loss function, based on a high-confidence pseudo-label generation and self-distillation iterative mechanism.

2. The method according to claim 1, characterized in that, The process of fusing the classification features and the shared features, and weighting the classification task to obtain the state probability of the target electrical appliance includes: Using a preset first fusion method, the classification features and the shared features are combined to form a classification-side joint representation, resulting in a classification fusion feature; The classification fusion features are weighted by acquiring and utilizing classification weights to obtain a weighted classification representation; Based on the weighted classification representation, output the state probability of the target electrical appliance.

3. The method according to claim 2, characterized in that, The process of fusing the regression features and the shared features, and weighting the regression task to obtain the power estimation data for the target electrical appliance includes: Using a preset second fusion method, the regression features and the shared features are combined to form a regression-side joint representation, resulting in regression fusion features; The regression fusion features are weighted by acquiring and utilizing the regression weights to obtain a weighted regression representation; Based on the weighted regression representation, output the power estimation data of the target electrical appliance.

4. The method according to claim 3, characterized in that, The step of generating and outputting the power decomposition result of the target electrical appliance based on the state probability and the power estimation data includes: The power decomposition result of the target electrical appliance is obtained by multiplying the state probability and the power estimation data by element-wise multiplication.

5. The method according to claim 4, characterized in that, The training steps for the non-invasive load monitoring model include: Acquire and use historical aggregate power sequences as input samples; Acquire and construct a multi-source weak supervision tag set based on the multi-source weak supervision information of the target electrical appliance; Construct an initial monitoring model and a joint loss function with weak supervision; Using the weakly supervised joint loss function as a constraint, the initial monitoring model is trained using the input samples and the multi-source weakly supervised label set to obtain the target monitoring model; A first target sample is obtained, and the first target sample is input into the target monitoring model to obtain the target output result. Then, a pseudo-label sample is generated using the sample data corresponding to the target output result that meets the preset high confidence condition. Obtain the original weakly supervised samples, merge the original weakly supervised samples with the pseudo-labeled samples to obtain the second target samples, and use the second target samples to train the target monitoring model to obtain the non-intrusive load monitoring model.

6. The method according to claim 5, characterized in that, The aggregated power sequence includes multiple power segments with preset sliding window lengths; The step of inputting the aggregated power sequence into the non-intrusive load monitoring model to obtain the power decomposition result of the target electrical appliance includes: Each of the power segments is sequentially input into the non-intrusive load monitoring model to obtain the power decomposition results of the target electrical appliance. After inputting the aggregated power sequence into the non-intrusive load monitoring model to obtain the power decomposition result of the target electrical appliance, the process further includes: The power decomposition results are reconstructed to obtain the continuous power decomposition results of the target electrical appliance.

7. The method according to claim 6, characterized in that, The weakly supervised joint loss function includes at least two of the following: aggregation consistency loss, state loss, over-limit penalty, non-negativity constraint, rated power ratio constraint, and event consistency loss.

8. A non-invasive load monitoring device, characterized in that, include: The acquisition module is used to acquire the aggregated power sequence to which the target electrical appliance belongs; The monitoring module is used to input the aggregated power sequence into a pre-trained non-intrusive load monitoring model to obtain the power decomposition results of the target electrical appliance; The non-intrusive load monitoring model includes a classification branch, a shared branch, a regression branch, and a gated fusion layer. The shared branch extracts shared features from the aggregated power sequence. The classification branch extracts classification features from the aggregated power sequence, fuses the classification features and the shared features, and performs a classification task weighting to obtain the state probability of the target appliance. The regression branch extracts regression features from the aggregated power sequence, fuses the regression features and the shared features, and performs a regression task weighting to obtain the power estimation data of the target appliance. The gated fusion layer generates and outputs the power decomposition result of the target appliance based on the state probability and the power estimation data. The non-intrusive load monitoring model is trained using the multi-source weak supervision label set of the target electrical appliance and the weak supervision joint loss function, based on a high-confidence pseudo-label generation and self-distillation iterative mechanism.

9. An electronic device, characterized in that, The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the method as described in any one of claims 1-7 according to instructions in the program code.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code for performing the method as described in any one of claims 1-7.

Images