Non-intrusive target load identification method based on cloud-end scene migration

Abstract

The invention relates to the technical field of power system monitoring and load management, and particularly discloses a non-intrusive target load identification method based on cloud-end scene migration, which introduces a physical consistency data enhancement mechanism at a cloud end, and simulates a modulation effect of voltage sag on a background load by calculating a dynamic coupling modulation factor. And a high-fidelity enhanced triple data set is generated, so that the problem that the coupling characteristic of a circuit is neglected in traditional data enhancement is solved. On this basis, a brand invariance feature extractor is trained by using hierarchical metric learning, and class essential features are learned by optimizing an inter-class and inter-brand distance forced model. Finally, the feature extractor and the classification head are assembled into a generalization model to be deployed to the end side, reasoning is carried out by freezing parameters of the feature extractor, the cross-brand generalization problem caused by physical heterogeneity of the electric appliance is effectively solved, and accurate identification of the electric appliance with no brand at the end side is achieved.

Description

Technical Field

[0001] This application relates to the field of power system monitoring and load management technology, and more specifically, to a non-intrusive target load identification method based on cloud-to-device scenario migration. Background Technology

[0002] With the advancement of the energy internet construction, non-intrusive load monitoring technology has become a key means of obtaining refined load data at the user-side equipment level due to its advantages of not requiring the installation of sensors deep inside the user's premises, low cost, and simple construction. This technology analyzes voltage and current data at the user's main incoming line to decompose and identify the operating status and energy consumption details of each internal electrical device, providing crucial data support for power load resource assessment, energy efficiency management, and demand response.

[0003] While deep learning-based load identification algorithms have achieved some success on specific datasets, in real-world cloud-edge collaborative applications, the heterogeneity of electrical physical implementations—that is, differences in internal components, control strategies, and circuit designs among appliances of the same type but different brands or models—leads to drastically different electrical characteristics on the bus circuit. Existing technologies largely rely on data-driven deep learning models. These models, adhering to the closed-world assumption, cannot correctly map the features of new brands of appliances with significant characteristic differences that have never been seen in the field to the known category space, resulting in identification failure and limiting the model's transferability and generalization performance across different user scenarios. Furthermore, to compensate for the scarcity of training data on the edge, data augmentation techniques are typically used to expand the sample. However, existing background noise superposition methods are mostly based on the simplified assumption that signal and noise are linearly independent, simply treating background noise as additive interference for synthesis. This approach ignores the physical coupling relationships that exist in actual circuits; that is, the occurrence of high-power electrical events can cause voltage dips in local power grid branches, thereby dynamically modulating the electrical behavior of background loads on those branches. Because the dynamic modulation effect of background noise driven by the main event could not be simulated, the physical fidelity of the synthetic data was insufficient, and it could not truly reproduce the coupling phenomenon in complex electrical environments. This further limited the robustness and recognition accuracy of the model trained using such data when facing real complex working conditions.

[0004] Therefore, an optimized, non-intrusive target load identification scheme based on cloud-to-edge scenario migration is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a non-intrusive target load identification method based on cloud-to-edge scenario migration, comprising:

[0006] Obtain the original standard dataset, where each record contains high-frequency current waveform data and its corresponding metadata, including appliance type and brand / model;

[0007] Physical consistency data augmentation and triple construction are performed on the original standard dataset to obtain an augmented triple dataset.

[0008] Based on the enhanced triplet dataset, the brand invariant feature extractor is trained to obtain the trained brand invariant feature extractor.

[0009] The model parameters of the trained brand invariant feature extractor are frozen, and the classification head is trained using a cloud-labeled dataset to obtain a trained classification head. The trained brand invariant feature extractor and the trained classification head together form a cloud-based generalization model.

[0010] Deploy the cloud-based generalization model on the edge-side load identification terminal;

[0011] The acquired electrical event waveforms are input into a cloud-based generalization model deployed on the load identification terminal at the edge to obtain predicted electrical type labels.

[0012] Compared to existing technologies, this method first constructs a raw dataset containing a rich variety of appliance types and brands in the cloud, and then introduces a physical consistency data augmentation mechanism. By extracting the instantaneous power envelope of events and calculating the dynamic coupling modulation factor, it simulates the voltage sag modulation effect of high-power events on the background load, generating an augmented triplet dataset with high physical fidelity, thus solving the problem of traditional data augmentation neglecting circuit coupling characteristics. Based on this, a brand-invariant feature extractor is trained using hierarchical metric learning. By optimizing inter-class and inter-brand distances, the model is forced to learn the essential features of appliance categories rather than brand-specific surface features. Finally, the trained feature extractor and classification head are assembled into a generalization model and deployed to the edge. At the edge, the feature extractor parameters are frozen, and only the classification head is fine-tuned or used for inference. This effectively overcomes the cross-brand generalization problem caused by the heterogeneity of appliance physical implementation, achieving accurate identification of appliances from brands not seen at the edge. Attached Figure Description

[0013] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0014] Figure 1This is a flowchart of a non-intrusive target load identification method based on cloud-to-device scene migration according to an embodiment of this application;

[0015] Figure 2 This is a data flow diagram illustrating the non-intrusive target load identification method based on cloud-edge scene migration according to an embodiment of this application;

[0016] Figure 3 This is a flowchart illustrating the process of performing physical consistency data augmentation and triple construction on the original standard dataset to obtain an enhanced triple dataset using a non-intrusive target load identification method based on cloud-edge scene migration according to an embodiment of this application.

[0017] Figure 4 This is a flowchart illustrating the process of performing physical consistency data augmentation on a clean event waveform set to obtain an enhanced event waveform set based on a background noise library in the non-intrusive target load identification method based on cloud-edge scene migration according to an embodiment of this application.

[0018] Figure 5 This is a flowchart illustrating the process of randomly extracting background noise waveforms from a background noise library and superimposing them with the time-domain transformed waveforms to obtain the final enhanced waveform in the non-intrusive target load identification method based on cloud-edge scene migration according to an embodiment of this application.

[0019] Figure 6 This is a flowchart illustrating the process of training a brand invariant feature extractor on an enhanced triplet dataset for a non-intrusive target load identification method based on cloud-to-device scene migration, according to an embodiment of this application, to obtain a trained brand invariant feature extractor. Detailed Implementation

[0020] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0021] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0022] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0023] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0024] Current non-intrusive load monitoring technologies face a severe challenge in cross-brand generalization during practical applications. The heterogeneity in the physical implementation of appliances from different brands leads to significant differences in electrical characteristics, making it difficult for existing models relying on surface statistical features to identify new brands of appliances not yet seen on the edge. Simultaneously, traditional data augmentation methods often simplify the linear relationship between signal and background noise, neglecting the dynamic modulation effect of voltage sags caused by high-power events on the background load in actual circuits. This results in insufficient physical fidelity of the training data, further limiting the robustness of the model. Therefore, this application proposes a non-intrusive target load identification method based on cloud-edge scene migration. This method achieves efficient migration from general knowledge in the cloud to personalized scenarios on the edge by constructing a physically consistent data augmentation mechanism and a brand-invariant feature extraction framework. Specifically, this method first introduces a physically consistent data augmentation mechanism in the cloud. By extracting the instantaneous power envelope of events and calculating a dynamic coupling modulation factor based on the hyperbolic tangent function, it accurately simulates the voltage sag modulation effect of high-power events on the background load, generating an augmented triplet dataset with high physical fidelity, thereby solving the problem of traditional data augmentation neglecting circuit coupling characteristics. Building upon this foundation, a brand-invariant feature extractor is trained using hierarchical metric learning. By constructing triples and optimizing the hierarchical triple loss function, combined with hard negative sample mining techniques, the model is forced to ignore superficial brand differences and focus on learning the essential physical characteristics of appliance categories. Finally, the trained feature extractor parameters are frozen and assembled with a classification head to form a generalized model, which is then deployed to the edge device. This allows edge devices to directly utilize the domain-invariant features learned in the cloud for inference, effectively overcoming the cross-brand generalization problem caused by the physical heterogeneity of appliances.

[0025] Figure 1 This is a flowchart of a non-intrusive target load identification method based on cloud-to-device scene migration according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating the non-intrusive target load identification method based on cloud-edge scenario migration according to an embodiment of this application. Figure 1 and Figure 2 As shown, the non-intrusive target load identification method based on cloud-edge scene migration according to an embodiment of this application includes: S100, acquiring an original standard dataset, wherein each record in the original standard dataset contains high-frequency current waveform data and its corresponding metadata, the metadata including appliance type and brand / model; S200, performing physical consistency data augmentation and triple construction on the original standard dataset to obtain an enhanced triple dataset; S300, training a brand invariant feature extractor based on the enhanced triple dataset to obtain a trained brand invariant feature extractor; S400, freezing the model parameters of the trained brand invariant feature extractor, and training a classification head through a cloud-labeled dataset to obtain a trained classification head, wherein the trained brand invariant feature extractor and the trained classification head constitute a cloud generalization model; S500, deploying the cloud generalization model on the edge-side load identification terminal; S600, inputting the acquired appliance event waveform into the cloud generalization model deployed on the edge-side load identification terminal to obtain the predicted appliance type label.

[0026] Specifically, in step S100, an original standard dataset is obtained. Each record in the original standard dataset contains high-frequency current waveform data and its corresponding metadata, including appliance type and brand / model. It is understood that, due to the challenges of cross-brand and cross-model generalization in practical applications, single or insufficient training data can lead to the model overfitting to the electrical characteristics of a specific brand, failing to grasp the essential physical laws of appliance categories, and thus failing to identify devices not previously encountered at the end-user level. Therefore, in the technical solution of this application, an original standard dataset is obtained, where each record contains high-frequency current waveform data and its corresponding metadata, including appliance type and brand / model. This constructs a data foundation with rich physical heterogeneity, clearly distinguishing the subtle differences and common features of different brands of appliances in terms of electrical characteristics. This provides accurate supervision signals for subsequent construction of physically consistent data enhancement and hierarchical metric learning, ensuring that the model can learn the essential characteristics of appliance categories independent of brand differences, thereby improving the model's robustness and generalization ability in different user scenarios.

[0027] More specifically, in a concrete example of this application, the construction of the original standard dataset follows a rigorous data acquisition and annotation process. First, a high-frequency data acquisition device is deployed in an experimental environment or real-world user scenario with diverse electrical appliance configurations. The sampling frequency of this device is set to a level sufficient to capture the transient characteristics of the appliances, such as kilohertz to tens of kilohertz, to ensure complete recording of detailed current waveform information. Next, controlled appliance operation tests are performed, sequentially starting and stopping target appliances of different brands and models, such as fixed-frequency and variable-frequency air conditioners of different brands, and water pumps or motors of different power specifications, while simultaneously recording the current waveform data at the main input line. Subsequently, the acquired raw waveform data is finely annotated, not only marking the start and stop times of the appliances but also recording detailed appliance category attributes, such as air conditioners, refrigerators, and washing machines, as well as specific brand and model information, forming a one-to-one corresponding data record. Finally, these records containing high-frequency current waveform segments and their detailed metadata were compiled into a structured database to form the original standard dataset. Each record is clearly associated with a specific electrical behavior and its physical entity identity, providing detailed data support for subsequent analysis of the characteristic differences of different brands of electrical appliances during startup transient and steady-state operation.

[0028] Specifically, in step S200, the original standard dataset undergoes physical consistency data augmentation and triplet construction to obtain an enhanced triplet dataset. It is understood that due to the complex physical coupling effects in actual power lines, the startup of high-power appliances is not merely a simple superposition of currents; it also triggers voltage dips at the point of common coupling, dynamically modulating the electrical behavior of background loads on the same line. This results in samples generated by traditional additive noise augmentation methods lacking physical fidelity, and training samples from a single brand are insufficient to cover the diverse equipment models and operating conditions at the end-point. Therefore, in the technical solution of this application, the original standard dataset is further augmented with physical consistency data and triplet construction to obtain an enhanced triplet dataset. This introduces dynamic disturbances conforming to the physical laws of circuits during the data synthesis stage, accurately simulating the nonlinear modulation effect of voltage dips on background noise, and constructing a comparative sample structure for metric learning. This improves the physical realism of the synthesized training data, forcing the model to focus on the essential electrical characteristics of appliance categories rather than environmental noise or brand-specific surface features during subsequent training, thereby effectively enhancing the model's generalization and recognition capabilities when facing unknown electromagnetic environments and new brand equipment.

[0029] Figure 3 This is a flowchart illustrating the process of performing physical consistency data augmentation and triplet construction on the original standard dataset to obtain an augmented triplet dataset, according to the non-intrusive target load identification method based on cloud-edge scene migration in this application. Figure 3As shown, step S200 includes: S210, performing data cleaning and event extraction on the original standard dataset to obtain a clean event waveform set; S220, performing physical consistency data augmentation on the clean event waveform set based on the background noise library to obtain an enhanced event waveform set; S230, performing anchor point selection, positive sample selection, and negative sample selection from the enhanced event waveform set to obtain an enhanced triplet dataset.

[0030] In step S210, the original standard dataset is cleaned and events are extracted to obtain a clean event waveform set. It is understood that the original high-frequency current data inevitably contains sensor noise, dead pixels caused by communication anomalies, and a large amount of meaningless steady-state background data. Directly using such raw data for subsequent processing would introduce interference features and reduce the model's learning efficiency of the electrical behavior of the appliance. Therefore, in the technical solution of this application, the original standard dataset is further cleaned and events are extracted to obtain a clean event waveform set. This allows for the accurate extraction of effective segments containing the transient processes of appliance start-up and shutdown from the continuous time series, and the removal of outliers and high-frequency noise interference from the signal. This provides high-quality reference waveforms for subsequent physical consistency data enhancement, ensuring that the synthesized data retains the inherent characteristics of the appliance while possessing high physical authenticity.

[0031] More specifically, in a concrete example of this application, the execution process begins with a traversal scan of the original labeled dataset. Based on the pre-labeled event timestamps in each record, the center moment of the target electrical event is determined. Using this moment as a reference, a time window covering the steady-state, transient changes before the event, and the new steady-state after the event is set, and original event waveform segments are extracted from the continuous high-frequency current data stream. Subsequently, a rigorous data cleaning operation is performed on the extracted waveform segments to check for outliers or null values ​​caused by acquisition equipment failures. These abnormal data points are smoothed and corrected using local median filtering or nearest-neighbor interpolation algorithms. Simultaneously, a low-pass digital filter is applied to filter the waveform, removing high-frequency white noise beyond the normal operating frequency range of the electrical appliances, thereby obtaining a clean event waveform with a high signal-to-noise ratio. Finally, the clean waveform is associated and encapsulated with its corresponding appliance type, brand model, and event type label, and stored in a clean event waveform set as the basic material for subsequently generating physically consistent enhanced samples.

[0032] In step S220, physical consistency data augmentation is performed on the clean event waveform set based on the background noise library to obtain an enhanced event waveform set. It is understood that in a real power supply network, the startup of high-power appliances is not an isolated current superposition process; the instantaneous large current generated flows through the power supply line with internal resistance, causing an observable voltage dip at the point of common coupling. This voltage change immediately affects other background loads on the same line, causing passive dynamic suppression or fluctuations in the background load current. Traditional linear superposition augmentation methods ignore this physical coupling effect, resulting in insufficient physical fidelity of the synthesized data. Therefore, in the technical solution of this application, physical consistency data augmentation is further performed on the clean event waveform set based on the background noise library to obtain an enhanced event waveform set. This introduces a dynamic physical coupling mechanism between the main event and the background load during the data synthesis process, accurately simulating the dynamic modulation effect of background noise driven by the power fluctuation of the main event. This generates synthesized samples with high physical fidelity, enabling the trained model to maintain extremely high robustness in the face of waveform distortion caused by voltage fluctuations in real-world scenarios.

[0033] Figure 4 This is a flowchart illustrating the process of performing physical consistency data augmentation on a clean event waveform set to obtain an enhanced event waveform set, based on a background noise library, according to an embodiment of the non-intrusive target load identification method based on cloud-to-edge scene migration in this application. Figure 4 As shown, step S220 includes: S221, extracting a first clean event waveform from the clean event waveform set; S222, performing random amplitude fine-tuning on the first clean event waveform to obtain an amplitude-fine-tuned waveform; S223, performing transient interval time-domain transformation on the amplitude-fine-tuned waveform to obtain a time-domain transformed waveform; S224, randomly extracting a background noise waveform from the background noise library and superimposing it with the time-domain transformed waveform to obtain the final enhanced waveform.

[0034] In step S221, a first clean event waveform is extracted from the clean event waveform set. It is understood that training a highly robust load identification model requires massive amounts of sample data covering various operating conditions, while the actual number of clean samples collected and cleaned is often limited. Directly using these limited samples for training can easily lead to model overfitting, making it unable to cope with complex situations such as voltage fluctuations and load changes in real-world scenarios. Therefore, in the technical solution of this application, a first clean event waveform is further extracted from the clean event waveform set. This serves as the seed for each preprocessed real electrical waveform to enhance physical consistency data, initiating subsequent signal simulation and synthesis processes. This ensures that all generated enhanced data retains the core electrical characteristics of the original electrical appliance, providing the model with training materials that possess both physical realism and statistical diversity.

[0035] More specifically, in a concrete example of this application, the extraction process reads data objects from a pre-built set of clean event waveforms through traversal or random indexing. First, based on a set enhancement factor parameter, an iterative loop is initialized. In each iteration, a specific clean event record is located from the set using a unique index identifier. Next, the high-frequency current time-series data stored in this record, i.e., the first clean event waveform, is read, and the associated metadata, including the appliance type label and brand / model information, is unpacked and obtained. This extraction operation ensures that subsequent enhancement algorithms, such as amplitude fine-tuning, time-domain transformation, and noise superposition, operate on a clear, clean, and correctly labeled physical entity signal, thereby guaranteeing the logical rigor and physical effectiveness of the data enhancement process.

[0036] In step S222, the waveform of the first clean event is randomly amplitude-fine-tuned to obtain the amplitude-fine-tuned waveform. It is understood that during the actual operation of a power distribution network, due to the influence of line impedance voltage drop and dynamic changes in grid load, the voltage at the point of common coupling on the user side is not constant at its rated value, but rather fluctuates randomly within the allowable range. This voltage fluctuation directly causes corresponding expansion and contraction in the operating current amplitude of connected electrical appliances. If the model only learns the characteristics under standard voltage, it is highly susceptible to misjudgment when the voltage deviates. Therefore, in the technical solution of this application, the waveform of the first clean event is further randomly amplitude-fine-tuned to obtain the amplitude-fine-tuned waveform. This introduces a random disturbance factor simulating grid voltage fluctuations during the data augmentation stage, enabling the synthesized sample to cover the operating characteristics of electrical appliances under different voltage levels. This enhances the adaptability of the load identification model to current amplitude changes caused by voltage fluctuations, ensuring stable identification performance even in a real and complex power grid environment.

[0037] More specifically, in a concrete example of this application, the random amplitude fine-tuning process first sets a range parameter characterizing the normal voltage fluctuation range based on power grid power quality standards or actual observation data, for example, setting it to 0.95 to 1.05 times the rated voltage. Then, using a computer's pseudo-random number generation algorithm, a scalar value is independently extracted from this preset uniform distribution range as the scaling factor for this enhancement operation. Finally, the extracted first clean event waveform is treated as a high-dimensional time series vector, and a point-by-point multiplication operation between the vector and the scalar is performed, that is, each sampling point value in the waveform sequence is multiplied by the scaling factor, thereby obtaining the waveform data after the overall amplitude has undergone a linear scaling transformation. This mathematical transformation process effectively simulates the output response of the appliance under different voltage conditions while preserving the key morphological features of the appliance's startup waveform (such as slope, overshoot, and steady-state ripple), enriching the diversity of the training samples in the amplitude dimension.

[0038] In step S223, the waveform after amplitude fine-tuning is subjected to transient interval time-domain transformation to obtain the time-domain transformed waveform. It is understood that, due to differences in mechanical load, wear and aging of internal transmission components, and manufacturing tolerances, the transient duration of the same type of electrical equipment during startup is not constant but exhibits significant time drift. If this nonlinear change in the time domain is not learned by the model, it will lead to a decrease in the accuracy of identifying the same type of equipment under different operating conditions. Therefore, in the technical solution of this application, the waveform after amplitude fine-tuning is further subjected to transient interval time-domain transformation to obtain the time-domain transformed waveform. This allows for random nonlinear time-base scaling of the transient process of electrical appliance startup while maintaining the overall waveform topology and energy change trend, simulating the startup characteristics of the equipment under different load resistance or aging conditions. This expands the distribution range of training samples in the time dimension, forcing the deep learning model to decouple the transient features from the absolute time length, thereby improving its robustness when facing varying startup speeds of electrical appliances in real-world scenarios.

[0039] More specifically, in a specific example of this application, the transient interval time-domain transformation process first calculates the current envelope sequence (used as a power surrogate when no voltage is sampled) and its first-order difference sequence based on the current waveform after amplitude fine-tuning. A preset power change rate threshold is used to automatically locate the start and end times of drastic power changes, thereby accurately locking down the transient interval containing rich feature information. Next, several uniformly distributed anchor points are selected as control points on the time axis of this transient interval, and random small displacement perturbations are applied to the time coordinates of these control points according to a preset time jitter ratio parameter, such as 10%. Subsequently, a monotonically increasing nonlinear mapping curve, such as a cubic B-spline curve, is generated using these perturbed control points. This curve defines the correspondence between the original time axis and the transformed time axis. Finally, the waveform data within the transient interval is resampled and interpolated based on this nonlinear mapping relationship, while keeping the steady-state waveform time axis before and after the transient interval unchanged, thereby generating time-domain transformed waveforms with different start-up durations.

[0040] In step S224, background noise waveforms are randomly selected from the background noise library and superimposed on the waveform after time-domain transformation to obtain the final enhanced waveform. It is understandable that existing background noise superposition methods suffer from a deep technical problem. This stems from an oversimplified physical assumption that the startup or operation event (signal) of the target appliance is linearly independent of the background load (noise) generated by other operating appliances in the same circuit. Therefore, traditional mechanisms only perform data synthesis using a simple additive noise model. However, this approach ignores a crucial physical coupling relationship: the occurrence of a high-power electrical event triggers a momentary current drawdown on its local power grid branch, leading to an observable voltage dip. This voltage change driven by the main event directly affects and dynamically modulates all background loads on that branch, causing corresponding changes in their electrical behavior. For example, in a typical industrial production scenario, when a high-power stamping machine starts up, the resulting voltage drop on the bus will instantly reduce the current amplitude of the lighting system or heating equipment operating on the same line. In other words, the current of a purely resistive background load will decrease proportionally with the voltage sag. Because traditional methods fail to simulate this event-driven dynamic modulation effect of background noise, the synthetic data generated has inherent defects in physical fidelity and fails to fully reproduce the complex electrical coupling phenomena in the real world. This limits the robustness and generalization ability of models trained using such data.

[0041] Therefore, in the technical solution of this application, background noise waveforms are randomly extracted from the background noise library and superimposed on the waveform after time-domain transformation to obtain the final enhanced waveform. This implements a dynamic modulation noise superposition method based on event coupling. In this process, background noise is no longer regarded as a static and unchanging signal, but as a moduloable object dynamically affected by the instantaneous power of the main event. By calculating the instantaneous power envelope of the target event and combining it with the circuit coupling coefficient, the superposition weight of the background noise is dynamically adjusted so that it exhibits a suppression phenomenon that conforms to physical laws at the peak power of the target event. In this way, synthetic waveforms with higher physical fidelity can be generated, allowing the model to be pre-adapted to the voltage coupling interference commonly present in the real power grid environment during the training phase. Thus, during actual deployment, the influence of background noise can be accurately removed, and the target load characteristics can be precisely locked.

[0042] Figure 5 This is a flowchart illustrating the process of randomly extracting background noise waveforms from a background noise library and superimposing them onto a time-domain transformed waveform to obtain the final enhanced waveform, according to an embodiment of the non-intrusive target load identification method based on cloud-edge scene migration in this application. Figure 5As shown, step S224 includes: S2241, extracting the instantaneous power envelope of the waveform after time-domain transformation to obtain a power envelope sequence; S2242, determining the dynamic coupling modulation factor sequence based on the power envelope sequence; S2243, dynamically adjusting the noise superposition of the background noise waveform and the waveform after time-domain transformation based on the dynamic coupling modulation factor sequence and the background noise weight to obtain the final enhanced waveform.

[0043] In step S2241, the instantaneous power envelope of the event is extracted from the time-domain transformed waveform to obtain a power envelope sequence. It is understood that, due to the extraction of the instantaneous power envelope, in simulating the physical process of a high-power electrical event causing a local grid voltage dip and thus modulating background noise, it is necessary to first quantize a physical proxy quantity from the event waveform that can characterize its instantaneous load intensity. The original current waveform is a sinusoidal oscillating signal, and the alternation of its instantaneous positive and negative values ​​cannot be directly used as an indicator of the degree of grid capacity occupancy; the trend of its amplitude change must be extracted to reflect the real energy impact. Therefore, in the technical solution of this application, the instantaneous power envelope of the event is further extracted from the time-domain transformed waveform to obtain a power envelope sequence, thereby using signal processing techniques to demodulate an envelope time sequence representing the instantaneous power consumption level of the event from the oscillating waveform. This power envelope directly reflects the intensity of the electrical event's demand on the power grid at each moment. Therefore, establishing this intensity information is crucial for calculating the extent of its impact on background noise, providing an accurate and physically meaningful input for subsequent dynamic modulation calculations. This ensures that subsequent steps can accurately calculate the depth of the voltage sag and the suppression ratio of the background load based on this intensity.

[0044] More specifically, in a particular example of this application, the implementation process primarily relies on the Hilbert transform as a mathematical tool to perform envelope detection of the signal. First, the time-domain transformed waveform generated in the preceding steps is used as the input signal, which preserves the transient time-domain characteristics of the appliance's startup process. Then, the process transforms the input time-domain transformed waveform... The algorithm applies the Hilbert transform to obtain the analytic signal and then calculates the magnitude of this analytic signal. Specifically, the algorithm first performs a Hilbert transform on the input waveform to construct an imaginary part signal orthogonal to the original signal. The real part (the original signal) is then combined with the imaginary part to form a complex analytic signal. The magnitude of this complex number in the complex plane, i.e., the instantaneous amplitude, is then calculated, resulting in an envelope time series representing the instantaneous power consumption level of the event. For example, when faced with a typical motor starting event, the extracted power envelope sequence will exhibit a unipolar curve that rises sharply and then decays slowly, clearly depicting the energy demand change of the motor from rest to rated speed, completely eliminating interference from the 50Hz power frequency carrier. This step can be expressed by the formula:

[0045]

[0046] in, Sampling points of the waveform after time-domain transformation: the input event waveform after primary transformation at time... The amplitude of the sampling points. Hilbert transform operator: A mathematical operator that performs a Hilbert transform on an input time series. Power envelope sequence: representing the event waveform at time [time]. The envelope value of the instantaneous power or current consumption intensity. This implementation converts the AC current signal into a DC envelope signal characterizing the load intensity, laying a solid mathematical foundation for the subsequent construction of the dynamic coupling modulation factor.

[0047] In step S2242, a dynamic coupling modulation factor sequence is determined based on the power envelope sequence. It is understood that due to the nonlinear effect of voltage sag in real circuits, the voltage drop amplitude and load current (or power) do not always maintain a simple linear proportional relationship. When the event power is small, its impact on the grid voltage is also small, exhibiting essentially linear growth; however, when the event power increases and exceeds a certain threshold, its impact on the voltage gradually tends to saturate, rather than growing infinitely linearly. This physical characteristic stems from the capacity limitations of the power grid and the magnetic saturation characteristics of transformers. Therefore, in the technical solution of this application, a dynamic coupling modulation factor sequence is further determined based on the power envelope sequence to accurately simulate the nonlinear effect of voltage sag in real circuits. By introducing a nonlinear activation function with saturation characteristics, the power envelope is mapped to a suppression coefficient for background noise. This generates a time-varying modulation factor that accurately reflects the physical coupling strength, thereby providing an accurate weight sequence for subsequent dynamic noise superposition and ensuring that the synthesized data truly reflects the degree of suppression of background load under high-power events.

[0048] More specifically, in a particular example of this application, the implementation process utilizes the hyperbolic tangent function to construct a mapping relationship. Specifically, based on the event power envelope obtained in the previous step... A time-dependent dynamic modulation factor sequence is calculated through a nonlinear mapping that introduces a hyperbolic tangent function. This ingeniously simulates the physical characteristics of voltage sags: the hyperbolic tangent function. The S-shaped curve accurately reflects this initial sensitivity followed by saturation, making it closer to physical reality than a simple linear model. For example, for the startup process of a high-power motor rapidly increasing from zero to full load, the calculated modulation factor... It decreases rapidly with increasing power (indicating background noise suppression) and remains at a stable low value when the power reaches its peak, without oscillating violently due to small power fluctuations. Furthermore, a factor called the circuit coupling coefficient is introduced in the formula. Configurable hyperparameters, which represent the rigidity or internal resistance of a specific local circuit in a physical scenario. A larger one... A value of 0 indicates a softer circuit, where a high-power event can easily trigger a significant voltage sag; conversely, a smaller value indicates a softer circuit. A value of 0 indicates that the circuit is very robust. The introduction of this parameter enables the data augmentation mechanism itself to simulate different electrical environments. This step can be expressed by the formula:

[0049]

[0050] in, Given a power envelope sequence, the events are at time... The power envelope value. This is the maximum value of the power envelope during the entire event, used to normalize the power. is the circuit coupling coefficient, a configurable hyperparameter that characterizes the internal resistance or rigidity of a circuit. It is the hyperbolic tangent function, used as an activation function for nonlinear mapping. It is a dynamically coupled modulation factor sequence, representing the time... The modulation factor, which determines the degree to which background noise should be suppressed, typically ranges from approximately [value missing]. This calculation process ensures that the noise superposition weight at each moment strictly follows the physical laws of the circuit.

[0051] In step S2243, based on the dynamic coupling modulation factor sequence and background noise weights, the background noise waveform and the waveform after time-domain transformation are dynamically adjusted and superimposed to obtain the final enhanced waveform. It is understood that the fundamental purpose of this preferred mechanism is to improve the physical fidelity of the synthetic training data used to train the non-intrusive load identification model. However, in practical applications, the waveforms generated by traditional superposition methods often exhibit a simple linear superposition of signal and noise, failing to reflect the nonlinear modulation phenomena commonly present in real-world scenarios. Therefore, in the technical solution of this application, the background noise waveform and the waveform after time-domain transformation are further dynamically adjusted and superimposed based on the dynamic coupling modulation factor sequence and background noise weights to obtain the final enhanced waveform. This performs dynamic modulation noise superposition, organically combining the event signal, background noise signal, and dynamic modulation relationship. By accurately simulating the dynamic modulation and suppression effects of high-power electrical events on the background load in the same circuit, the data enhancement samples generated by this method can more realistically reproduce the complex coupling phenomena in the actual electrical environment. The load identification model trained using this high-fidelity data possesses stronger robustness and better generalization performance. When the model is deployed in diverse and electromagnetically complex real-world scenarios, it is better able to understand and respond to waveform distortions caused by physical coupling, thereby maintaining higher recognition accuracy when faced with unseen combinations of equipment and operating conditions.

[0052] More specifically, in a particular example of this application, the implementation process achieves point-by-point fusion of waveforms through mathematical operations. Specifically, the randomly selected background noise waveform... With a base noise weight used to control the overall signal-to-noise ratio Multiply, and then combine with the dynamically coupled modulation factor sequence generated in the previous step. The noise signal is obtained by multiplying it point by point. Then, this modulated noise signal is compared with the input event waveform. By adding the values ​​point by point, the final enhanced waveform is obtained. This step fully simulates the entire physical coupling effect. Unlike traditional mechanisms, the background noise term here... It is no longer static; its amplitude will change with the instantaneous power of the main event (via...). This can manifest as dynamic fluctuations. For example, in a scenario involving the start-up and continuous operation of a high-power air conditioning compressor and background lighting load, the power of the event will peak (compressor start-up shock). Approaching a minimum, background noise (lighting current) is suppressed to the greatest extent; when the event ends, the power drops (the compressor enters steady state). Returning to 1, the background noise also returns to normal. Thus, this high-fidelity physical simulation achieves the effect of producing a synthetic waveform sample that highly matches the coupling characteristics of real electrical measurement data. This step can be expressed by the formula:

[0053]

[0054] in, These are the sampling points of the waveform after time-domain transformation; the input event waveform at time... The sampling points. The base noise weight is a random scalar value used to control the overall signal-to-noise ratio. For a dynamically coupled modulation factor sequence, at time... The dynamic modulation factor value. Background noise waveform sampling points, randomly selected background noise waveform at time... The sampling points. To enhance the final waveform sampling points, dynamic modulation noise was superimposed, resulting in the final enhanced waveform output at time [time value missing]. The sampling points. This process ensures that each generated augmented sample contains real physical interaction logic.

[0055] In step S230, anchor point selection, positive sample selection, and negative sample selection are performed from the enhanced event waveform set to obtain the enhanced triplet dataset. It is understandable that due to the heterogeneity in physical structure and control logic of different brands of electrical appliances, the electrical features of the same category are discretely distributed in the feature space, and even the intra-class differences are greater than the inter-class differences, making it difficult for traditional supervised learning models to form compact category clusters. Therefore, in the technical solution of this application, anchor point selection, positive sample selection, and negative sample selection are further performed from the enhanced event waveform set to obtain the enhanced triplet dataset. This explicitly constructs the constraint relationship of attraction between different brands of the same category and repulsion between different categories at the data level. In particular, by selecting samples from different brands as positive samples, a strong cross-brand feature alignment signal is provided for the subsequent model. This forces the deep learning model to strip away non-essential features related to brands during training and focus on extracting invariant features that can characterize the commonalities of electrical appliance categories, thereby fundamentally improving the model's generalization accuracy for electrical appliances of unknown brands.

[0056] More specifically, in a concrete example of this application, the execution process first establishes a multi-level index structure based on the set of enhanced event waveforms. This index uses appliance type and brand / model as keys to facilitate quick retrieval of samples with specific attributes. Then, an iterative process for generating triples is initiated. In each iteration, an enhanced waveform is randomly selected from the set as an anchor sample, and its appliance type and brand label are read. Next, based on a brand generalization strategy, a waveform with the same appliance type identifier as the anchor sample but belonging to a different brand or model is prioritized and selected as a positive sample. If the dataset lacks samples from different brands, waveforms from the same original event source but with different random enhancement processes are selected as substitutes. Simultaneously, a waveform with an appliance type completely different from the anchor sample is retrieved and selected as a negative sample. Finally, the selected anchor, positive, and negative samples are encapsulated into an independent triple data unit, and this process is repeated until an enhanced triple dataset meeting a preset size is constructed.

[0057] Specifically, in step S300, the brand-invariant feature extractor is trained based on the enhanced triplet dataset to obtain a trained brand-invariant feature extractor. It is understandable that in practical cross-domain applications of non-intrusive load identification, the heterogeneity of the physical implementations of different brands of electrical appliances leads to extremely discrete distributions of the same type of electrical features in the feature space. There are even instances where the feature distance between different brands within the same category is much greater than the feature distance between different categories, making it difficult for traditional classification models to define clear decision boundaries in the feature space, thus failing to identify brand equipment not previously seen at the end-user. Therefore, in the technical solution of this application, the brand-invariant feature extractor is further trained based on the enhanced triplet dataset to obtain a trained brand-invariant feature extractor. This utilizes a deep metric learning mechanism to optimize the triplet loss function, forcing the model to narrow the distance between samples of different brands within the same category during feature mapping, while simultaneously widening the distance between samples of different categories, thereby reconstructing the topology of the feature space. This forces the feature extractor to strip away non-essential interfering features related to brand and model, focusing instead on extracting common, invariant features that characterize the working principle and physical properties of appliances. This ensures that when faced with appliances from entirely new brands, the model can still accurately map their feature vectors to the correct category cluster, achieving cross-brand generalization.

[0058] Figure 6 To obtain a trained brand-invariant feature extractor based on the enhanced triplet dataset, the non-intrusive target load identification method based on cloud-to-device scene migration according to the embodiments of this application is used. Figure 6As shown, step S300 includes: S310, inputting the enhanced triplet dataset into the brand invariant feature extractor to obtain batches of anchor feature vectors, positive sample feature vectors, and negative sample feature vectors; S320, performing hard negative sample mining on the batches of anchor feature vectors and negative sample feature vectors to obtain batches of hard negative sample feature vectors; S330, constructing a hierarchical triple loss function based on the batches of hard negative sample feature vectors, anchor feature vectors, positive sample feature vectors, and negative sample feature vectors; S340, performing backpropagation of gradient descent on the brand invariant feature extractor based on the hierarchical triple loss function to obtain the trained brand invariant feature extractor.

[0059] In step S310, the enhanced triplet dataset is input into the brand-invariant feature extractor to obtain batches of anchor feature vectors, positive sample feature vectors, and negative sample feature vectors. It is understandable that, since the original enhanced triplet waveform data is a high-dimensional time-series signal, directly measuring distance in the original signal space is easily affected by phase deviations, amplitude fluctuations, and non-critical high-frequency noise, making it difficult to accurately quantify the semantic similarity between different samples. Therefore, in the technical solution of this application, the enhanced triplet dataset is further input into the brand-invariant feature extractor to obtain batches of anchor feature vectors, positive sample feature vectors, and negative sample feature vectors. This utilizes the nonlinear mapping capability of deep neural networks to project the high-dimensional current waveform into a low-dimensional and compact feature embedding space. This compresses key information representing the electrical attributes of electrical appliances into feature vectors, providing a mathematical representation basis for subsequent calculation of the relative distance between samples in the feature space and the construction of the metric loss function.

[0060] More specifically, in a concrete example of this application, the process first selects a certain number of triplet samples sequentially or randomly from the enhanced triplet dataset according to a preset batch size parameter, forming an input data batch containing anchor points, positive samples, and negative samples. Next, the waveform data in this batch is converted into a tensor format adapted to the network input and fed in parallel into the input layer of a brand-invariant feature extractor, which typically employs a shared-weight Siamese neural network architecture or a multi-branch convolutional structure. Subsequently, through hierarchical transfer between network layers, a series of convolutional operations are performed to capture local transient changes in the waveform. Combined with nonlinear activation functions and pooling operations, a global steady-state contour is extracted, gradually filtering out background interference from the original waveform and extracting high-order abstract semantic features. Finally, at the network's output layer, the deeply encoded multidimensional feature map is mapped into fixed-dimensional real-valued vectors, thereby outputting the anchor point feature vector batch, positive sample feature vector batch, and negative sample feature vector batch, each corresponding to one of the input batches.

[0061] In step S320, hard negative sample mining is performed on the anchor point feature vector batch and the negative sample feature vector batch to obtain the hard negative sample feature vector batch. It is understandable that, since most negative samples in the randomly sampled training batch are naturally far from the anchor point samples in the feature space, they are easily distinguishable simple negative samples. If the loss is calculated based solely on these samples, the gradient value generated by the model will approach zero, making it difficult for the network to learn the crucial ability to distinguish easily confused electrical appliances with highly similar electrical features but different categories. Therefore, in the technical solution of this application, hard negative sample mining is further performed on the anchor point feature vector batch and the negative sample feature vector batch to obtain the hard negative sample feature vector batch. This allows for real-time calculation of the distance matrix between the anchor point and all candidate negative samples within the feature space, and for each anchor point, the closest and most difficult-to-distinguish negative sample vector is accurately located and extracted. This ensures that the training process always focuses on solving the most ambiguous classification boundary in the feature space, forcing the model to uncover the deep essential features that can effectively distinguish similar loads (such as different types of motor equipment), thereby improving the discrimination accuracy and robustness of the feature extractor in complex and ever-changing load scenarios.

[0062] More specifically, in a specific example of this application, hard negative sample mining is performed on the anchor feature vector batch and the negative sample feature vector batch to obtain the hard negative sample feature vector batch, including: calculating the distance matrix between the anchor feature vector batch and the negative sample feature vector batch; performing hard negative sample index positioning based on the distance matrix to obtain the hard negative sample index vector; and extracting the hard negative sample feature vector batch from the negative sample feature vector batch based on the hard negative sample index vector.

[0063] Accordingly, the distance matrix between the batch of anchor feature vectors and the batch of negative sample feature vectors is calculated, and hard negative sample indexing is performed based on the distance matrix to obtain the hard negative sample index vector. It is understandable that, in the process of constructing the appliance feature space based on metric learning, randomly sampled negative samples (such as microwave ovens) often differ significantly from anchor samples (such as air conditioners) in electrical features, resulting in a naturally large distance between them in the feature space. The loss value generated by such simple negative samples is negligible and cannot provide an effective gradient direction for optimizing model parameters, making it difficult for the model to learn the crucial ability to distinguish easily confused appliances (such as specific models of water pumps and air conditioners) with highly similar electrical features but different categories. Therefore, in the technical solution of this application, the distance matrix between the batch of anchor feature vectors and the batch of negative sample feature vectors is further calculated, and hard negative sample indexing is performed based on the distance matrix to obtain the hard negative sample index vector. This allows for a panoramic quantification of the similarity between each anchor sample and all potential negative samples in the feature space within the current training batch, and accurately identifies the most challenging and difficult-to-distinguish negative sample positions for the current model. This ensures that the subsequent training process always focuses on solving the problem of the most ambiguous classification boundary in the feature space, forcing the model to dig out the deep essential features that can effectively distinguish similar loads, thereby improving the discrimination accuracy of the feature extractor in complex and ever-changing load scenarios.

[0064] More specifically, in a concrete example of this application, the implementation process involves mathematical operations performed immediately after the feature extractor outputs feature vectors. First, batches of anchor feature vectors and negative sample feature vectors output by the brand-invariant feature extractor are obtained. These batches contain high-dimensional waveform feature representations mapped by a deep neural network. Next, matrix operations are used to compute the squared pairwise Euclidean distances between these two vector sets in parallel, constructing a distance matrix with dimensions equal to the batch size multiplied by the batch size. Each element in this matrix precisely represents the distance between the corresponding anchor appliance and the negative sample appliance in the current feature space. Subsequently, a minimum index search operation is performed on each row of this distance matrix; that is, for each anchor, the column index of the element with the smallest value in the corresponding row vector is found. This column index directly points to the negative sample in the current batch that is closest to that specific anchor, i.e., the one with the most similar features. Finally, the minimum distance indices corresponding to all anchors are aggregated to generate a hard negative sample index vector, which will serve as the addressing basis for accurately extracting hard negative sample feature vectors from the subsequent negative sample batches.

[0065] Accordingly, the hard negative sample feature vector batch is extracted from the batch of negative sample feature vectors based on the hard negative sample index vector. It is understood that since the subsequent calculation of the hierarchical triple loss function relies on the specific coordinates of the feature vectors in high-dimensional space to quantify the semantic distance between samples, simply knowing the index position of the hard negative samples is insufficient for gradient backpropagation; the feature entities of these key samples must be extracted for computation. Therefore, in the technical solution of this application, the hard negative sample feature vector batch is further extracted from the batch of negative sample feature vectors based on the hard negative sample index vector, thereby physically retrieving the high-dimensional feature representations of those negative samples most easily confused with the anchor samples from the current device batch based on the positioning results. This allows the construction of a high-difficulty negative sample set that strictly corresponds one-to-one with the anchor samples, providing precise vector input for the subsequent calculation of the hard negative sample boundary loss, thus forcing the model to optimize the subtle boundaries of the feature space.

[0066] More specifically, in a concrete example of this application, the extraction process is implemented using a vector indexing mechanism based on tensor operations. First, the hard negative sample index vector generated in the preceding steps is received, which stores the index of the minimum distance negative sample corresponding to each anchor sample. Then, using this index vector as a lookup table, batch aggregation operations are performed on the original negative sample feature vectors. Specifically, for the first... The element value is directly accessed, and the corresponding feature vector in the negative sample batch is copied to the new tensor. Each position. Through this parallelized addressing and data reassembly process, a brand new batch of hard negative sample feature vectors is reassembled. Although this batch has the same size as the anchor batch, each element within it is the most difficult obstacle for the current model to distinguish, thus preparing the data for calculating the boundary constraints between similar but different brands and the most similar outliers.

[0067] In step S330, a hierarchical triple loss function is constructed based on batches of hard negative sample feature vectors, batches of anchor point feature vectors, batches of positive sample feature vectors, and batches of negative sample feature vectors. It is understandable that in the feature space of electrical loads, appliances of the same type but different brands may have large intra-class distances due to differences in physical implementation, while appliances of certain different types may have small inter-class distances due to similar electrical characteristics. A single metric learning objective often struggles to simultaneously consider global category classification and local boundary subdivision, making it difficult for the model to distinguish between target loads of a specific brand and highly similar interference loads. Therefore, in the technical solution of this application, a hierarchical triple loss function is further constructed based on batches of hard negative sample feature vectors, batches of anchor point feature vectors, batches of positive sample feature vectors, and batches of negative sample feature vectors. This establishes a dual geometric constraint mechanism in the feature space, including coarse-grained class intervals and fine-grained hard example intervals. This forces the model to compress the distance between samples of the same type but different brands while focusing on pushing the most easily confused outlier samples outside the safe boundary, thereby improving the model's identification accuracy under complex interference.

[0068] More specifically, in a concrete example of this application, the construction process of the function involves a weighted combination and threshold constraints of multiple sets of feature distances. First, for each anchor sample, its Euclidean distance in the feature space with positive samples (different brands within the same category), ordinary negative samples, and hard negative samples (most similar outliers) is calculated. Next, these distances are used to construct a first-level category loss term, requiring that the distance between the anchor and ordinary negative samples is at least a larger first preset interval than the distance between the anchor and positive samples, in order to maintain the separability of the basic categories. Simultaneously, a second-level hard negative sample loss term is constructed, requiring that the distance between the anchor and hard negative samples is at least a smaller second preset interval than the distance between the anchor and positive samples, in order to handle confusion at the boundaries. Finally, these two parts of the loss are truncated and summed using a maximum value function to form the final scalar loss value, which will serve as the direct basis for subsequent gradient calculations, driving the network parameters to continuously optimize towards intra-class compactness and inter-class dispersion.

[0069] In step S340, the brand-invariant feature extractor is backpropagated using gradient descent based on the hierarchical triple loss function to obtain the trained brand-invariant feature extractor. More specifically, in a concrete example of this application, this implementation follows a standard deep learning optimization process. First, an optimizer instance, such as an adaptive moment estimation optimizer, is initialized and bound to all trainable parameters of the brand-invariant feature extractor. In each training iteration, the cached gradient information from the previous round in the optimizer is first cleared, and then the backpropagation interface of the hierarchical triple loss value is called. This operation triggers an automatic differentiation engine, which backpropagates the error layer by layer from the loss function layer to the input layer according to the chain rule, accurately calculating the partial derivative of the total loss with respect to the weights and biases of each convolutional kernel in the network. Finally, the optimizer performs a parameter update step according to the preset learning rate strategy and the calculated gradient direction, fine-tuning the millions of parameters in the feature extractor, and repeating this process until the model's loss converges on the validation set.

[0070] Understandably, the calculated hierarchical triple loss value only quantifies the topological deviation of the current model in the feature space mapping—that is, the similar samples are not compact enough or the hard negative samples are not far enough apart—without directly changing the internal logic of the model in processing waveform data. Without parameter updates, the model cannot master how to extract the correct features. Therefore, in the technical solution of this application, based on the hierarchical triple loss function, gradient descent backpropagation is performed on the brand-invariant feature extractor to obtain a trained brand-invariant feature extractor. This transforms the abstract geometric constraint objective into specific network parameter adjustment instructions, driving the model to internalize the ability to process complex electrical waveforms through continuous trial and error and correction. In this way, the identification logic of "ignoring brand differences and focusing on the essence of the category" can be solidified in the weight structure of the neural network, thereby obtaining a feature extractor with strong cross-brand generalization ability.

[0071] Specifically, in step S400, the model parameters of the trained brand-invariant feature extractor are frozen, and the classification head is trained using a cloud-labeled dataset to obtain a trained classification head. The trained brand-invariant feature extractor and the trained classification head together form a cloud-based generalization model. It is understood that since the brand-invariant feature extractor, after training with hierarchical metric learning, has mastered the ability to map the original current waveform to a category common feature space stripped of brand differences, further adjustments to its parameters during the subsequent construction of the classification decision logic can easily disrupt the established robust feature topology, causing the model to refit to non-essential details of specific brands in the training data, thereby weakening cross-brand generalization performance. Therefore, in the technical solution of this application, the model parameters of the trained brand-invariant feature extractor are further frozen, and the classification head is trained using a cloud-labeled dataset to obtain a trained classification head. The trained brand-invariant feature extractor and the trained classification head together form a cloud-based generalization model. This approach maintains the feature extraction network's immunity to brand differences while utilizing supervised learning to establish an efficient mapping relationship from abstract feature vectors to specific appliance category labels. In this way, a complete model can be assembled that has both deep physical feature understanding and accurate classification decision-making capabilities, providing a powerful pre-training foundation for edge devices to directly deal with unseen branded appliances, while significantly reducing the computational complexity of subsequent edge fine-tuning.

[0072] More specifically, in a concrete example of this application, the implementation process first sets the weight parameters of all convolutional, pooling, and normalization layers in the brand-invariant feature extractor to an untrainable state within the deep learning framework, i.e., blocking the gradient update path for these layers during backpropagation. Next, a lightweight classification head structure is constructed at the output of the feature extractor, typically consisting of one or two fully connected layers and a softmax activation function at the end, used to compress the high-dimensional feature vectors and map them to probability distributions corresponding to various appliances (such as air conditioners, refrigerators, and washing machines). Subsequently, using the full sample data from the cloud-labeled dataset, a standard supervised training process is executed: waveform data is input into the frozen feature extractor to generate stable feature vectors, and these vectors are then input into the classification head to calculate the cross-entropy loss between the predicted probability and the true class label. Finally, based on this loss value, gradient descent optimization is performed only on the connection weights of the classification head until the classification accuracy reaches a preset standard. The solidified feature extractor and the trained classification head are then packaged into a unified cloud-based generalization model file, ready to be distributed to edge terminals.

[0073] Specifically, in step S500, the cloud-based generalization model is deployed on the load identification terminal at the edge. It is understandable that while cloud training relies on massive amounts of data to build a general model with strong cross-brand generalization capabilities, actual load identification services have extremely high real-time requirements. Furthermore, continuously uploading high-frequency current waveforms collected at the edge to the cloud for inference not only consumes huge amounts of network bandwidth but also involves security risks related to user energy privacy data. Therefore, in the technical solution of this application, the cloud-based generalization model is further deployed on the load identification terminal at the edge, thereby sinking the prior knowledge accumulated in the cloud, including the ability to extract brand-invariant features, to the edge computing node closer to the physical location of the user's load. This enables the edge device to still have the ability to instantly identify appliances of unknown brands even offline or in weak network environments, achieving closed-loop data processing locally. This ensures both response speed and data privacy, while also laying the foundation for subsequent online adaptive fine-tuning of the model using local data.

[0074] More specifically, in a concrete example of this application, the deployment process encompasses three key stages: lightweight model conversion, secure delivery, and local runtime loading. First, on a cloud server, the trained cloud-based generalized model (including a parameter-frozen feature extractor and an initial classification head) is serialized and converted into a lightweight inference format adapted to embedded hardware architectures (such as ARM processors) to reduce model file size and memory usage. Next, via an encrypted remote communication channel (such as a 4G / 5G private network or encrypted fiber optic network), the model file is securely pushed to the load identification terminal's memory installed in an industrial park or commercial building's electrical distribution box using a remote wireless upgrade mechanism. Finally, the embedded operating system inside the terminal starts the loading program, reads the model parameters into memory, initializes the inference engine, and simultaneously establishes a direct data mapping channel between the model input layer and the terminal's underlying high-frequency analog-to-digital conversion driver, thus completing the physical migration from cloud-based algorithms to on-site computing power.

[0075] Specifically, in step S600, the acquired electrical event waveform is input into the cloud-based generalization model deployed on the load identification terminal at the end side to obtain the predicted electrical type label. It is understandable that, due to the large number and rapid updates of electrical equipment brands in the actual operating environment at the end side, new models or brands of equipment not included in the cloud training set often appear. Although these devices have the same physical functions, their electrical characteristics differ. If the model lacks the ability to generalize and infer unknown brands, it will lead to missed or false alarms. Therefore, in the technical solution of this application, the acquired electrical event waveform is further input into the cloud-based generalization model deployed on the load identification terminal at the end side to obtain the predicted electrical type label. This utilizes the brand-invariant feature extraction logic embedded in the model to map the unfamiliar waveform captured at the end side to known category feature clusters, and performs inductive reasoning based on the learned category commonalities. This enables accurate identification of electrical appliances of brands not seen on-site, breaking the limitation of traditional data-driven models that can only identify known samples, and ensuring that the load monitoring system can still output reliable identification results when facing an open and dynamic equipment environment.

[0076] More specifically, in a concrete example of this application, the inference process is a pipelined operation driven by real-time data streams. First, the high-frequency acquisition thread inside the load identification terminal continuously monitors changes in the total incoming current. Once the instantaneous power change rate exceeds a preset threshold, an event capture mechanism is immediately triggered, capturing a waveform segment of the electrical event containing the complete transient process. Next, the waveform segment undergoes preprocessing operations consistent with those used during cloud training, including normalization and pixelation format conversion, to adapt it to the model's input interface standard. Subsequently, the processed waveform data tensor is input into the cloud-based generalization model loaded in the terminal's memory. The model's feature extractor first converts it into a high-dimensional feature vector insensitive to brand differences. Then, the classification head calculates the probability distribution of each electrical appliance category based on this feature vector. Finally, by comparing the probability values, the category with the highest probability is selected as the prediction result, i.e., the electrical appliance type label. This identification result, along with the event timestamp, is reported to the data management center through the terminal's communication interface, completing a full end-side identification task.

[0077] In summary, the non-intrusive target load identification method based on cloud-edge scene migration according to the embodiments of this application is clarified. First, a raw dataset containing rich appliance types and brand models is constructed in the cloud, and a physical consistency data augmentation mechanism is introduced. By extracting the instantaneous power envelope of events and calculating the dynamic coupling modulation factor, the voltage sag modulation effect of high-power events on the background load is simulated, generating an enhanced triplet dataset with high physical fidelity, thereby solving the problem of traditional data augmentation neglecting circuit coupling characteristics. Based on this, a brand-invariant feature extractor is trained using hierarchical metric learning. By optimizing inter-class distance and inter-brand distance, the model is forced to learn the essential features of appliance categories rather than the surface features unique to brands. Finally, the trained feature extractor and classification head are assembled into a generalization model and deployed to the edge. The feature extractor parameters are frozen on the edge, and only the classification head is fine-tuned or used for inference, effectively overcoming the cross-brand generalization problem caused by the heterogeneity of appliance physical implementation, and achieving accurate identification of appliances of brands not seen on the edge.

[0078] As described above, the non-intrusive target load identification method based on cloud-edge scene migration according to the embodiments of this application can be implemented in various power monitoring and metering devices, such as smart IoT energy meters, load identification terminals, new power load management systems, or smart park energy management platforms. In one possible implementation, the non-intrusive target load identification method based on cloud-edge scene migration according to the embodiments of this application can be integrated as a software module or hardware module into the power grid company's electricity consumption information collection system or cloud-edge-device collaborative load monitoring system. For example, this non-intrusive target load identification method based on cloud-edge scenario migration can be a load identification model pre-training and distribution application running on a cloud server, or it can be a device-level energy consumption breakdown and resource potential assessment function plug-in module of an existing electricity information collection master station, or it can be a middleware service deployed on an edge gateway and reporting identification results to the master station through an interface. Of course, the core physical consistency data enhancement, brand invariance feature extraction and edge model adaptive fine-tuning logic in this method can also be embedded in hardware such as smart meter modules, energy controllers, edge computing nodes or dedicated load identification terminal communication boards, as one of the device-level load real-time monitoring or refined energy efficiency management hardware modules of the user-side perception system.

[0079] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A non-intrusive target load identification method based on cloud-edge scenario migration, characterized in that, The method comprises the following steps: obtaining an original standard data set, each record in the original standard data set containing high-frequency current waveform data and corresponding metadata, the metadata including appliance type and brand / model; performing physical consistency data augmentation and triple construction on the original standard data set to obtain an augmented triple data set; training a brand-invariance feature extractor based on the augmented triple data set to obtain a trained brand-invariance feature extractor; freezing the model parameters of the trained brand-invariance feature extractor, and training a classification head through a cloud labeling data set to obtain a trained classification head, wherein the trained brand-invariance feature extractor and the trained classification head constitute a cloud generalization model; deploying the cloud generalization model on an end-side load identification terminal; inputting the obtained appliance event waveform into the cloud generalization model deployed on the end-side load identification terminal to obtain a predicted appliance type label. 2.The cloud-end scene migration based non-intrusive target load identification method according to claim 1, characterized in that, The method of performing physical consistency data augmentation and triple construction on the original standard data set to obtain an augmented triple data set comprises the following steps: performing data cleaning and event extraction on the original standard data set to obtain a clean event waveform set; performing physical consistency data augmentation on the clean event waveform set based on a background noise library to obtain an augmented event waveform set; selecting anchor points, positive samples and negative samples from the augmented event waveform set to obtain an augmented triple data set. 3.The cloud-end scenario migration based non-intrusive target load identification method according to claim 2, characterized in that, The method of performing physical consistency data augmentation on the clean event waveform set based on a background noise library to obtain an augmented event waveform set comprises the following steps: extracting a first clean event waveform from the clean event waveform set; performing random amplitude fine-tuning on the first clean event waveform to obtain an amplitude fine-tuned waveform; performing transient interval time domain transformation on the amplitude fine-tuned waveform to obtain a time domain transformed waveform; randomly extracting a background noise waveform from the background noise library and superimposing it on the time domain transformed waveform to obtain a final augmented waveform.

4. The cloud-end scene migration based non-intrusive target load identification method according to claim 3, characterized in that, The method of randomly extracting a background noise waveform from the background noise library and superimposing it on the time domain transformed waveform to obtain a final augmented waveform comprises the following steps: performing event transient power envelope extraction on the time domain transformed waveform to obtain a power envelope sequence; determining a dynamic coupling modulation factor sequence based on the power envelope sequence; performing dynamic adjustment noise superposition on the background noise waveform and the time domain transformed waveform based on the dynamic coupling modulation factor sequence and the background noise weight to obtain the final augmented waveform.

5. The cloud-end scene migration based non-intrusive target load identification method according to claim 4, characterized in that, The method of determining a dynamic coupling modulation factor sequence based on the power envelope sequence comprises the following steps: determining the dynamic coupling modulation factor sequence according to the following formula: wherein, is a power envelope sequence, is a maximum power envelope value; is a circuit coupling coefficient, is a hyperbolic tangent function, is a dynamic coupling modulation factor sequence. 6.The cloud-end scene migration based non-intrusive target load identification method of claim 1, wherein, The method of training a brand-invariance feature extractor based on the augmented triple data set to obtain a trained brand-invariance feature extractor comprises the following steps: inputting the augmented triple data set into the brand-invariance feature extractor to obtain an anchor point feature vector batch, a positive sample feature vector batch and a negative sample feature vector batch; performing hard negative sample mining on the anchor point feature vector batch and the negative sample feature vector batch to obtain a hard negative sample feature vector batch; Based on the hard negative sample feature vector batch, the anchor point feature vector batch, the positive sample feature vector batch and the negative sample feature vector batch, a hierarchical triple loss function is constructed; Based on the hierarchical triple loss function, the brand invariance feature extractor is subjected to gradient descent back propagation to obtain a trained brand invariance feature extractor.

7. The cloud-end scenario migration based non-intrusive target load identification method according to claim 6, characterized in that, Hard negative sample mining is performed on the anchor point feature vector batch and the negative sample feature vector batch to obtain the hard negative sample feature vector batch, including: A distance matrix between the anchor point feature vector batch and the negative sample feature vector batch is calculated; Based on the distance matrix, hard negative sample index positioning is performed to obtain a hard negative sample index vector; Based on the hard negative sample index vector, the hard negative sample feature vector batch is extracted from the negative sample feature vector batch.

Images