Low-frequency power load identification method and system based on event driving and layered identification

Abstract

The application provides a kind of event-driven and layered identification low-frequency power load identification method and system, it is related to power load monitoring technical field, wherein, the method comprises: based on power mutation detection, in the power data sampled at low frequency, the switching event of power load is identified, and the identification data window containing the switching event is intercepted;Load identification model is constructed, and self-supervised pre-training and fine-tuning are carried out on the load identification model;The identification data window of the load event to be identified is identified in layers using the load identification model;In the process of layered identification, events not belonging to known loads are open set rejection;Based on the result of open set rejection, the load identification model is continuously updated.Through the application, the fundamental defect that the performance of offline training model ages over time and cannot adapt to environmental changes is solved.

Description

Technical Field

[0001] This invention relates to the field of electrical load monitoring technology, and in particular to a method and system for identifying low-frequency electrical loads based on event-driven and hierarchical rejection. Background Technology

[0002] The development of smart grids has made data collection at the power user end more convenient. Non-Intrusive Load Monitoring (NILM) technology, in particular, identifies the operating status and energy consumption of each appliance by collecting low-frequency signals such as voltage and current at the main electricity meter. Most existing NILM methods rely on pattern recognition or optimization algorithms to decompose the electricity consumption information of individual appliances using the collected electricity data. However, these existing technologies have the following shortcomings in practical applications: 1. Reliance on high-frequency or large amounts of labeled data: Many methods require high-frequency sampled data or large amounts of manually labeled historical data to build a device feature template library. Obtaining high-frequency data or a comprehensive appliance database is costly and difficult in reality, and the recognition accuracy is often insufficient with low-frequency (e.g., 10Hz) data. Existing low-frequency solutions often require collecting a large amount of operating data for individual appliances to train the model, resulting in a large amount of modeling data, which is not conducive to rapid deployment.

[0003] 2. Inability to adapt to open-set scenarios (unknown devices): Traditional methods typically assume that the device set is fixed (closed set) during training and recognition. In real-world home or electrical environments, users may add new appliances or replace equipment; such unknown loads cannot be correctly identified by existing models. Existing models often misclassify when encountering unseen loads, lacking the ability to reject unknown devices.

[0004] 3. Lack of hierarchical identification and fine-grained classification: Existing methods mostly output device categories directly, making it difficult to distinguish different models or power levels of the same type of equipment. For load management and refined power consumption analysis, only broad categories (such as "air conditioners") are identified, but specific models or power levels cannot be identified, limiting the depth of application.

[0005] 4. Real-time performance and edge deployment limitations: Due to hardware resource constraints, edge devices (such as smart gateways) cannot run complex deep learning models for extended periods. Some algorithms have slow recognition speeds or require data to be uploaded to the cloud for processing, failing to meet real-time requirements. Furthermore, continuous operation of complex algorithms increases the computing power and energy consumption burden on edge devices, hindering stable operation in low-power environments.

[0006] 5. Difficulty in Continuously Optimizing the Model: Changes in the power environment over time, equipment aging, and the addition of new equipment all affect recognition accuracy. Most existing methods employ offline training and fixed model deployment, making it difficult to improve model performance in response to environmental changes. There is a lack of a continuous learning mechanism to allow the model to continuously absorb new data and optimize its recognition capabilities after deployment.

[0007] The aforementioned issues make it difficult for existing non-intrusive load identification solutions to fully meet the requirements in terms of low-frequency sampling, open-set identification, identification precision, and real-time edge applications. Therefore, it is necessary to provide a new technical solution that can utilize low-frequency sampling data to identify the type and model of power loads in real time and efficiently on edge devices, reject unknown loads, and support subsequent continuous learning and updates, thereby improving the practicality and intelligence of non-intrusive load monitoring. Summary of the Invention

[0008] This invention provides a low-frequency power load identification method and system based on event-driven and hierarchical rejection, which solves the defects of existing low-frequency power load identification methods in the face of unknown loads, such as insufficient identification ability, coarse identification granularity, and difficulty in adaptive model updating.

[0009] This invention provides a low-frequency power load identification method based on event-driven and hierarchical rejection, comprising: Based on power surge detection, switching events of electrical loads are identified in low-frequency sampled power data, and an identification data window containing the switching events is extracted. A load identification model is constructed, and the load identification model is subjected to self-supervised pre-training and fine-tuning. The load identification model is used to perform hierarchical identification of the identification data window of the load event to be identified; During the hierarchical identification process, events that do not belong to the known load are rejected using an open set method. Based on the results of the open set rejection, the load identification model is continuously learned and updated.

[0010] According to the present invention, a low-frequency power load identification method based on event-driven and hierarchical rejection is provided. Based on power surge detection, it identifies switching events of power loads in low-frequency sampled power data and extracts an identification data window containing the switching events, including: Voltage and current data of the power system bus are acquired using a low-frequency sampling frequency; Monitor changes in the voltage and current data, and identify switching events of electrical loads using a power surge detection algorithm; When the detected change in power or voltage exceeds a preset threshold, a switching event is determined to have occurred, and a time window centered on the moment the switching event occurs is extracted as an identification data window.

[0011] According to the low-frequency power load identification method based on event-driven and hierarchical rejection provided by the present invention, before constructing the load identification model, the method further includes: Extract multi-dimensional features reflecting the electrical characteristics of the electrical load from the identification data window; Multiple view samples are generated by applying a physically consistent data augmentation strategy to the original feature sequence, which consists of multiple features.

[0012] According to the present invention, a low-frequency power load identification method based on event-driven and hierarchical rejection is provided, which constructs a load identification model and performs self-supervised pre-training on the load identification model, including: Construct a feature encoding network for the load identification model; Multiple enhanced view samples of the same load event are input into the feature encoding network to obtain corresponding feature vectors; The feature encoding network is trained by a self-supervised contrastive learning algorithm, so that the feature vectors of the same load event are close to each other and the feature vectors of different load events are separated from each other. After pre-training, the feature encoding network is connected to at least one classifier to form a complete load recognition model.

[0013] According to the present invention, a low-frequency power load identification method based on event-driven and hierarchical rejection is provided, wherein the load identification model is fine-tuned, including: Collect tagged data for known load events; The labeled known load events are input into the load identification model, the error between the classifier output and the true category label is calculated, and the model parameters are updated so that the load identification model can distinguish between predefined load categories. The output layer of the load recognition model is also configured with a mechanism for identifying unknown categories.

[0014] According to the present invention, a low-frequency power load identification method based on event-driven and hierarchical rejection is provided, which uses the load identification model to perform hierarchical identification on the identification data window of the load event to be identified, including: Input the identification data window of the load event to be identified into the load identification model, and output the load category to which the load event to be identified belongs; For each load category to which the load event to be identified belongs, the corresponding load model or device identifier is further output.

[0015] According to the present invention, a low-frequency power load identification method based on event-driven and hierarchical rejection is provided. In the hierarchical identification process, open-set rejection is performed on events that do not belong to known loads, including: If the confidence level of the load identification model in the load event to be identified not belonging to any known category or model is lower than the threshold, a rejection decision is triggered, the load event to be identified is marked as an unknown load, and no identification result of the existing category is output; The rejection determination is based on a preset threshold rule or feature space distance discrimination to detect and reject unknown loads in open set scenarios.

[0016] According to the present invention, a low-frequency power load identification method based on event-driven and hierarchical rejection is provided, wherein the load identification model is continuously learned and updated based on the results of the open-set rejection, including: The data of the load events to be identified, which will be marked as unknown loads, will be added to the training dataset; Event samples marked as unknown loads are assigned new categories or new model labels, and the load recognition model is retrained or fine-tuned by incorporating previously trained data to update the model parameters to recognize unknown loads. The updated load identification model is deployed to edge devices for the identification of subsequent load events to be identified.

[0017] This invention also provides a low-frequency power load identification system based on event-driven and hierarchical rejection, comprising: An event detection module is used to identify switching events of electrical loads in low-frequency sampled power data based on power surge detection, and to capture an identification data window containing the switching events. The training and fine-tuning module is used to build a load identification model and perform self-supervised pre-training and fine-tuning on the load identification model. The load identification module is used to perform hierarchical identification of the identification data window of the load event to be identified using the load identification model; The rejection discrimination module is used to reject events that do not belong to the known load during the hierarchical identification process by performing open set rejection. The learning and updating module is used to continuously learn and update the load identification model based on the results of the open set rejection.

[0018] According to the present invention, a low-frequency power load identification system based on event-driven and hierarchical rejection is provided, the system comprising an edge device and a cloud server; The edge device includes the event detection module, the load identification module, and the rejection judgment module; The cloud server includes the training fine-tuning module and the learning update module; The training fine-tuning module and the learning update module are also used to distribute the load recognition model, which has been pre-trained by self-supervised training or continuously learned and updated, to the edge device.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. High-efficiency low-frequency identification with strong real-time performance: Employing event-driven window modeling significantly reduces data processing volume, it can still accurately capture key points of appliance start-up and shutdown even under low-frequency sampling conditions of 10Hz. The lightweight event detection algorithm and on-demand triggered identification process ensure stable long-term operation of edge devices, responding to each load event in real time and meeting the real-time requirements of millisecond-level event detection and second-level identification.

[0020] 2. Small modeling data volume and improved accuracy through self-supervision: By employing multi-view physical consistency enhancement and self-supervised pre-training strategies, this invention fully utilizes a large amount of unlabeled data to learn load feature representations, significantly reducing reliance on manually labeled data. Compared to traditional fully supervised learning, which requires collecting a large amount of switching data from various devices, this invention only needs a small number of known device samples to complete model fine-tuning, reducing model construction costs. Simultaneously, the representations obtained through self-supervised learning are more robust, improving the ability to distinguish between complex devices in multiple states and devices with similar features, resulting in higher recognition accuracy.

[0021] 3. Refined Hierarchical Recognition with Low Error Rate: This invention innovatively introduces a hierarchical recognition framework based on load category → model, which can further distinguish specific models or specifications after determining the major equipment categories, making the recognition results more informative. The hierarchical strategy reduces the confusion between different categories of equipment and improves model accuracy; for further subdivision within the same category, it also avoids the class imbalance problem that is prone to occur in general single-level classifiers. Therefore, the system can detect and identify multiple appliance categories and models at once, meeting the needs of refined energy consumption management.

[0022] 4. Open-set rejection ensures reliability: To address the continuous emergence of new devices in reality, this invention integrates open-set recognition and rejection mechanisms into the model. When encountering loads exceeding the training range, the system can intelligently detect and output an "unknown" category rather than an incorrect one, significantly reducing the false recognition rate. Compared to the information misleading caused by the blind judgment of traditional models, the rejection strategy improves the credibility of the recognition results and prepares for the inclusion of new devices.

[0023] 5. Continuous Learning and Adaptation: Utilizing a closed-loop mechanism of continuous learning triggered by rejection, the model can be continuously updated and expanded over time. Whenever a user adds a new appliance or the characteristics of existing equipment change, the system incorporates new knowledge into the model through incremental learning, achieving a transition from rejection to recognition of new loads. The model can continuously learn new load characteristics, keeping its recognition performance optimized with environmental changes, overcoming the shortcomings of traditional models that remain unchanged and gradually age, and significantly improving its practical application value.

[0024] 6. Edge-Cloud Collaboration and Optimized Resource Utilization: The two-stage deployment of this invention fully leverages the collaborative advantages of edge and cloud computing. Low computational overhead at the edge ensures real-time performance and user privacy; centralized training in the cloud enhances the global optimality of the model and can serve multiple nodes, exhibiting excellent scalability. Compared to pure cloud solutions, this invention reduces data transmission and cloud computing pressure; compared to pure local solutions, it overcomes edge computing power limitations, balancing real-time performance and model complexity. Therefore, the method of this invention possesses both economic efficiency and reliability for deployment in practical smart power consumption scenarios. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the low-frequency power load identification method based on event-driven and hierarchical rejection provided by the present invention; Figure 2 This is a schematic diagram illustrating the principle of event-driven window modeling in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hierarchical identification and rejection process in an embodiment of the present invention; Figure 4 This is a block diagram of the continuous learning closed-loop mechanism triggered by rejection in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the collaborative work between edge devices and the cloud under a two-stage deployment architecture in this embodiment of the invention; Figure 6 This is a structural block diagram of the low-frequency power load identification system based on event-driven and hierarchical rejection provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. 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.

[0028] Please see Figures 1 to 5This application provides a low-frequency power load identification method based on event-driven and hierarchical rejection, aiming to solve the problems of insufficient power load identification accuracy, inability to identify unknown loads, and difficulty in continuously updating the model under low-frequency sampling conditions. The core of this method lies in efficiently processing low-frequency sampling data through an event-driven approach, and combining hierarchical identification, open-set rejection, and continuous learning mechanisms to achieve accurate, fine-grained, and adaptive identification of power loads.

[0029] Figure 1 This is a flowchart of the low-frequency power load identification method based on event-driven and hierarchical rejection provided by the present invention, as shown below. Figure 1 As shown, this method includes the following steps: Step S101: Based on power mutation detection, identify the switching events of electrical loads in the power data sampled at low frequency, and extract the identification data window containing the switching events; Step S102: Construct a load identification model and perform self-supervised pre-training and fine-tuning on the load identification model; Step S103: Use the load identification model to perform hierarchical identification of the identification data window of the load event to be identified; Step S104: During the hierarchical identification process, open-set rejection is performed on events that do not belong to the known load. Step S105: Based on the results of the open set rejection, the load identification model is continuously learned and updated.

[0030] The method provided in this application first identifies switching events of electrical loads in low-frequency sampled power data based on power surge detection, and then extracts an identification data window containing these switching events. This step is the foundation of an event-driven mechanism, designed to transform continuous, massive data streams into discrete, valuable event fragments, thereby significantly reducing the computational burden of subsequent processing and enabling real-time analysis on resource-constrained edge devices. By triggering the identification process only when a significant power change is detected, invalid calculations for stable operating states are avoided, solving the problems of poor real-time performance and high energy consumption caused by continuous data processing in existing technologies.

[0031] After acquiring the identification data window, this method constructs a load identification model and performs self-supervised pre-training and fine-tuning on it. This step aims to address the challenge of traditional methods relying on large amounts of manually labeled data. Through self-supervised pre-training, the model can learn general electrical feature representations of loads from massive amounts of unlabeled data, significantly reducing the cost and time required for model construction. Subsequently, only a small amount of labeled data is needed for fine-tuning, enabling the model to quickly adapt to specific identification tasks. This training strategy effectively addresses the shortcomings of background techniques—large modeling data volume and difficult deployment—by fully utilizing unlabeled data, thereby improving the model's generalization ability and initial accuracy.

[0032] Next, the trained load recognition model is used to perform hierarchical recognition on the recognition data window of the load event to be identified. The purpose of hierarchical recognition is to improve the granularity and accuracy of the recognition. It decomposes the complex task of full-category recognition into two relatively simple sub-tasks: "first determine the broad category, then identify the model." This approach not only provides more valuable fine-grained recognition results (such as specific model numbers), but also reduces the probability of confusion between different categories by setting up specialized classifiers at different levels, thereby solving the shortcomings of the prior art in terms of coarse recognition granularity and difficulty in distinguishing similar devices.

[0033] During the hierarchical identification process, this method performs open-set rejection of events that do not belong to the known load. This is a crucial step in addressing the continuous emergence of new devices (i.e., open-set scenarios) in practical applications. Its design aims to endow the model with the ability to "know what it knows and admit what it doesn't know." When the model encounters a load it has never seen before, it refuses to give an incorrect classification result, instead labeling it as "unknown." Through this mechanism, the reliability of the identification results is significantly improved, decision-making errors caused by misclassification are avoided, and the serious problem in the background technology where models inevitably misjudge when facing unknown loads is solved.

[0034] Finally, based on the results of open-set rejection, the load identification model is continuously learned and updated. This step constructs a closed-loop system from discovering the unknown to learning the unknown. Its design aims to enable the model to have adaptive and self-evolving capabilities. The rejected unknown load data is collected and used for incremental training of the model, enabling the model to continuously learn and master newly emerging load characteristics. In this way, the model can keep pace with changes in the power environment, solving the fundamental defect of offline-trained models in the background technology that the performance of offline-trained models ages over time and cannot adapt to environmental changes.

[0035] Further, in a preferred embodiment, the specific process of identifying switching events of electrical loads in low-frequency sampled power data based on power surge detection, and extracting an identification data window containing the switching events, includes: First, acquiring voltage and current data of the power system bus at a low-frequency sampling frequency (e.g., in the range of 1Hz to 10Hz). Then, monitoring changes in the voltage and current data in real time, and using a power surge detection algorithm to identify switching events of electrical loads. When a change in power or voltage is detected to exceed a preset threshold, for example, a power change ΔP exceeding a threshold Th1 (where Th1 is a preset power change threshold), a switching event is determined to have occurred. Figure 2 As shown, once an event is determined to have occurred, a time window centered on the moment the switch event occurs is extracted as the identification data window. This window includes the steady-state interval before and after the event and the intermediate transitional change segment, providing complete information for subsequent feature extraction.

[0036] Specifically, a sliding window analysis (e.g., a window length of N points, centered on the current moment) is performed on continuous power data to calculate the power jump characteristics within the window. If a significant change in power is detected before and after the center of the window and the event determination condition is met, an appliance switching event is considered to have occurred at this moment. The determination condition may be: the power change ΔP exceeds a preset power change threshold Th1, and the steady-state mean before and after the change is significantly different, while also meeting the shape similarity index requirement to filter out short-term fluctuations not caused by device switching. Once the event is confirmed, the event timestamp is recorded, and the relevant time interval data is extracted, i.e., the voltage and current sequence extended forward and backward for a predetermined duration (e.g., from 2 seconds before to 5 seconds after) centered on the event point, forming an identification data window. During extraction, it is ensured that the beginning and end of the window are in a relatively steady state to avoid transitional noise. In this embodiment, for periods when no event is detected, the gateway only performs monitoring without triggering identification, and the window slides along the time axis to continue monitoring the next moment.

[0037] More specifically, due to the low sampling rate of 10Hz, power change detection needs to balance sensitivity to low-power devices and robustness to noise fluctuations. Therefore, this embodiment employs a sliding window detection algorithm with tiered decision-making: 1. Sliding Window Power Difference: Define a sliding window of length M, and calculate the average power difference ΔP between the second half and the first half of the window at each new sampling point. w If ΔP w Exceeding the first threshold Th p (For example, 20W, selected based on the power of typical small devices), then it is preliminarily considered that an event may occur. Simultaneously, the rate of change of variance of the power sequence within the window is calculated; if there is a sudden increase, it indicates a sudden change in the power curve. Meeting these two conditions triggers the candidate event.

[0038] 2. Similarity Coefficient Filtering of Steady-State Fluctuations: For candidate event points, the difference between short-term steady-state segments before and after the event (e.g., average power 0.5 seconds before and after the event point) and the waveform shape similarity are further calculated. If the difference exceeds the second threshold Th... s If the waveforms before and after the event have low similarity (indicating significant differences between the states), then it is confirmed as a genuine load switching event; otherwise, it is judged as a false event caused by fluctuations in the power of an existing operating load and is ignored. For example, the periodic start-stop of a refrigerator compressor or air conditioner compressor may cause brief power fluctuations, but these fluctuations are usually small in amplitude and short in duration. By setting thresholds and similarity conditions, they can be distinguished from genuine switching events.

[0039] 3. Event Deduplication: When an event is detected, the detection output is temporarily frozen for a short period (e.g., 1-2 seconds) to avoid multiple oscillations during the startup process of the same device being misjudged as multiple events. The sliding window can skip this frozen interval and continue moving. In addition, multiple events occurring simultaneously (in rare cases where two devices start up sequentially within a sampling interval) can be split or processed in combination based on the power change amplitude. Generally, simultaneous events are difficult to distinguish at a 10Hz sampling granularity, but this embodiment focuses on simplifying the process. That is, if a composite power jump is detected within a window, the main change can be located, and then the remaining changes can be detected in the next window.

[0040] Through the above multi-step judgment, interference from noise and large load fluctuations can be eliminated, accurately capturing the true start-up and shutdown points of the equipment. The thresholds and parameters set in this embodiment can be optimized according to the actual environment. For example, the variance increment threshold can be appropriately increased for environments with high noise levels, while the power difference threshold can be decreased for monitoring small electrical appliances. Simultaneously, this embodiment also monitors abnormal voltage changes: if a significant voltage drop occurs at a certain moment (e.g., >5% of the rated value), it may indicate the connection of a high-power load, which can help determine the occurrence of the event and avoid missed detections.

[0041] like Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the principle of event-driven window modeling in this embodiment of the invention. For example, in the user's total power-time curve, the sliding window captures the moment of power jump as the event point, and takes data of a certain time length before and after the event as the recognition window (shaded area). The figure shows the power jump at the center point of the window and the steady-state interval before and after the window, to illustrate how to exclude short-term fluctuations and accurately extract event features.

[0042] In another preferred embodiment, before constructing the load identification model, data preprocessing is performed to enhance the model's learning ability. Specifically, firstly, multi-dimensional features reflecting the electrical characteristics of the load are extracted from the identification data window. These features may include active power P, reactive power Q (calculated from voltage and current phasors), power factor PF (reflecting the current-voltage phase relationship), total harmonic distortion (THD) (estimated from current waveform sampling), and transient variation characteristics of voltage and current, etc. Then, a physically consistent data augmentation strategy is applied to the original feature sequence of the multi-dimensional features to generate multiple view samples. For example, a certain amplitude of noise perturbation, scaling, or phase shift is applied to the voltage and current sequences simultaneously to ensure that the resulting changes in features such as P, Q, and PF are consistent with real physical laws; or the original sequence is slightly shifted or subjected to local time distortion to simulate measurement errors or sample different starting points. Through these augmentations, multiple physically consistent "view" samples are obtained for self-supervised pre-training of the model. Physical consistency ensures that each augmented view still maintains the characteristic patterns of real load events and does not introduce pseudo-signals that violate energy conservation or circuit laws. This enhancement provides high-quality positive sample pairs for subsequent self-supervised learning, which helps the model learn more robust feature representations.

[0043] For example, a feature vector X is calculated from the data within the identification data window. Vector X includes, for example, the pre-event steady-state power P. prev Post-event steady-state power P post (The difference between the two, ΔP, represents the load's operating power), the corresponding reactive power difference, ΔQ, power factor change, ΔPF, and the change in current waveform distortion, ΔTHD, etc. Additionally, transient characteristics during the event can be included, such as current peaks and voltage dips (e.g., slight voltage drops may occur when certain devices start up). These multidimensional features ensure a comprehensive characterization of the load's electrical properties.

[0044] Physically consistent data augmentation strategies include at least one of the following operations: 1. The current signal is scaled up by a random ratio, and the calculated active and reactive power are adjusted accordingly to simulate the effect of power supply voltage changes. 2. Add small-amplitude random noise to the voltage and current signals, and adjust the calculation of power factor and harmonic distortion accordingly to reflect the physical characteristics under noise conditions. 3. Apply phase shift or time axis translation to voltage and current signals synchronously to ensure that the relative timing of voltage and current remains consistent, thereby maintaining the consistent power change pattern during load start-up and shutdown; 4. High-frequency components are superimposed on the current signal to increase the total harmonic distortion (THD), and the effective value of the current is adjusted accordingly to simulate the harmonic background under nonlinear load conditions.

[0045] The enhanced view samples described above all correspond to the same actual load event and are used as positive sample pairs during self-supervised comparative learning.

[0046] Furthermore, the process of constructing a load recognition model and performing self-supervised pre-training includes: First, constructing a feature encoding network, typically employing a Convolutional Neural Network (CNN) or Temporal Convolutional Network (TCN) structure to effectively process temporal data. Then, inputting multiple view samples generated from the same load event after data augmentation into the feature encoding network to obtain their respective feature vectors. The feature encoding network is trained using a self-supervised contrastive learning algorithm (such as InfoNCE loss), the core idea of ​​which is to make the feature vectors generated by different view samples of the same load event closer to each other in the feature space, while feature vectors from different load events are further apart. After pre-training, this feature encoding network, which already possesses good feature extraction capabilities, is connected to at least one classifier to form a complete load recognition model ready for fine-tuning.

[0047] Specifically, the load recognition model is self-supervised pre-trained using the enhanced multi-view data mentioned above. Self-supervised algorithms such as contrastive learning are introduced: two different enhanced views of the same event window are used as positive sample pairs, and data from different event windows are used as negative samples to train the load recognition model for representation learning. Specifically, a temporal feature extraction network (such as an encoder based on a convolutional neural network or temporal convolutional network) is constructed during the pre-training stage. The enhanced time series or feature sequences are input, and low-dimensional feature vectors are output. By maximizing the similarity of feature vectors between different views of the same event and minimizing the similarity of features from different events, the model learns a representation that reflects the inherent characteristics of the load. This self-supervised pre-training process does not require manual labeling and can be completed using a large amount of unlabeled data, significantly reducing the dependence on labeled samples. The pre-trained model has mastered a general feature representation of load events.

[0048] After obtaining the initial model through self-supervised learning, supervised fine-tuning training is performed using a small amount of labeled load event data. Samples from historically collected event windows are manually labeled with their corresponding load categories and models, serving as the training set. This set is used to fine-tune the pre-trained feature extraction network and connect it to the classifier layer. Supervised training with limited labeled samples optimizes the model's discriminative ability for the actual loads to be identified. Because the model already possesses good initial feature representations, the amount of labeled data required for fine-tuning is far less than training a model from scratch. This results in a smaller modeling dataset and faster training convergence, improving the model's practicality.

[0049] For example, two enhanced view samples of the same load event are input into the encoding network to obtain two feature vectors, which are designated as positive sample pairs. Feature vectors from different load events are selected as negative samples. A contrastive loss function is defined to narrow the feature distance between the positive sample pairs and widen the feature distance between the event and the negative samples. The encoding network parameters are iteratively optimized through gradient descent so that the features output by the encoding network exhibit clustered and separated distributions for different loads in a high-dimensional space. Preferably, the encoding network is constructed using a convolutional neural network, a temporal convolutional network, or a long short-term memory (LSTM) network to extract feature representations of voltage and current time-series data.

[0050] In one alternative implementation, the process of fine-tuning the load identification model includes: first, collecting a small amount of labeled data of known load events, which are relatively inexpensive to label manually. Then, these labeled known load events are input into a pre-trained load identification model. The classifier in the model provides a predicted class, and the model's parameters are updated by calculating the error between this prediction and the true class label (e.g., cross-entropy loss) and using a backpropagation algorithm. This process aims to enable the load identification model to accurately distinguish between predefined load categories. Notably, during model design, its output layer is also configured with mechanisms for unknown class discrimination, such as setting a confidence threshold or a dedicated "unknown" output node, laying the foundation for subsequent open-set rejection.

[0051] For example, the feature encoding network of the model is trained using a large amount of historically measured 10Hz voltage and current data (without labels). First, a set of window segments containing single events is automatically extracted from the historical data. This can be achieved by locating each switching event segment in the long-term data using an offline event detection algorithm and cropping it into samples. Simultaneously, some steady-state segments without events can be included for comparison, enabling the model to learn to distinguish between "changing" and "unchanging" patterns.

[0052] For each cut event segment, two random data augmentation views are generated. Augmentation operations include, for example, multiplying the entire current sequence by a random coefficient within the range of [0.9, 1.1], and correspondingly multiplying the voltage by another coefficient close to 1, to simulate the impact of different operating voltage levels on power; or adding independent small noises (such as Gaussian noise with an amplitude of 1% of the signal's maximum value) to the voltage and current respectively, corresponding to instrument measurement errors; or shifting the voltage phase by a few degrees, while adjusting the calculated active / reactive components accordingly to maintain the physical relationship. For total harmonic distortion (THD), THD can be increased by superimposing small high-frequency harmonic components onto the current sequence to simulate changes under nonlinear load conditions, while ensuring the correspondence with the power factor. All these augmented views should maintain macroscopic characteristics consistent with the original sequence (e.g., unchanged equipment rated power and power factor category), only introducing certain perturbations to make the model less sensitive to small changes and more focused on core patterns.

[0053] Two augmented views of the same event are input into a feature encoding network with shared parameters to obtain feature vectors z1 and z2. During training, a contrastive learning loss function is used to bring z1 and z2 closer together in the feature space while keeping them further away from feature vectors of other different events. Specifically, an InfoNCE loss can be used: z1 is used as the anchor, z2 as the positive sample, and features from several other events are randomly selected as negative samples. The similarity between the anchor and the positive sample must be higher than the similarity between the anchor and any negative sample. Through continuous training iterations, the feature encoding network gradually learns to map similar load events (even after different augmentations) to similar representations, while mapping different load events to distinct representations. This means that the model extracts key features reflecting the device's identity. For example, for multiple samples of the same refrigerator compressor starting (collected under different ambient voltages and background noise), the feature encoding network can extract their common current surge features and steady-state power features, clustering the representations together; while the electric kettle activation event forms another cluster. This representation learning does not require manual labeling of categories but forms a good feature space, laying the foundation for downstream classification.

[0054] After completing self-supervised pre-training, the feature encoding network is solidified and then fine-tuned using labeled data. In this embodiment, switching data of several typical devices in an experimental environment are collected as a label set, such as lighting fixtures (LED lights at several power levels), fans, rice cookers, air conditioners, etc., with several samples in each category. Using this data, supervised training is performed on the end of the pre-trained model: a fully connected layer is added as the category classification output, and cross-entropy loss is used to train the model to identify the correct device category; at the same time, in order to take into account the open set recognition capability, the model also outputs an "unknown" category node, and samples that do not belong to any category in the training set are labeled "unknown" during training (if such samples exist, they can come from fragments of other devices to simulate the unknown class). After a few rounds of fine-tuning training, the classification part of the model converges, and its overall architecture has the ability to discriminate categories and a certain degree of rejection capability.

[0055] For model-level segmentation, a finer-classification output layer can be added after the category classifier, and corresponding fine-classification training data can be prepared for each category. Since many model details may be difficult to distinguish in low-frequency data, fine-classification models can typically utilize richer features (such as differences in device rated power, startup duration, etc.). In this embodiment, a power threshold rule is introduced for some categories (such as LED lighting fixtures) to distinguish models: for example, 15W or less is classified as a light bulb, and more than 15W is classified as a ceiling light, etc., to assist the model in outputting specific model inferences. This rule can be considered a simple special case of the model to improve the reliability of fine-classification.

[0056] In this embodiment, by freezing most of the parameters of the encoding network and training only the output layer, or by training the entire network with a small learning rate, the model can adapt to the classification task of a specific load set while maintaining its original representational capabilities. Furthermore, the model is given open-set recognition capability during fine-tuning training, including: introducing an output node representing an "unknown" category during training; assigning unknown labels to samples that are not part of the training target category for training; or using a data hold-out strategy to create pseudo-unknown samples in the training data, thereby teaching the model to reduce output confidence when encountering non-training categories.

[0057] Furthermore, the specific process of using the load identification model to perform hierarchical identification of the identification data window of the load event to be identified is as follows: First, the event data to be identified is input into the model. The first-level classifier of the model identifies the load category and outputs the load category to which the event belongs, such as "lighting equipment" or "kitchen appliances". After determining the load category, if there are more specific models to be identified under this category, the second-level classifier of the model will further process them and output the corresponding load model or specific device identifier, for example, identifying "15W LED light" under the "lighting equipment" category. This coarse-to-fine hierarchical strategy effectively improves the accuracy and granularity of the identification.

[0058] Specifically, the load identification model in this embodiment consists of a two-level classifier: the first level is a category classifier, used to determine which category of electrical appliance the event belongs to; the second level is a model classifier, which further identifies the specific device model given the category. The category classifier can be a fully connected neural network, which outputs the probabilities of several predefined categories via softmax. In this embodiment, common household appliances are divided into several categories, such as: lighting, kitchen appliances, entertainment electronics, refrigeration equipment, heating equipment, and motors. When feature X is input into the category classifier, if the probability of a certain category (e.g., "kitchen appliances") is significantly higher than others and exceeds a confidence threshold, then the event is determined to belong to this category.

[0059] Subsequently, a model classifier corresponding to the "kitchen appliances" category is selected to perform more granular identification of feature X. This model classifier can be a sub-model specifically trained for kitchen appliances. For example, for the kitchen appliance category, the model identifies subcategories such as "microwave oven," "rice cooker," and "oven," and may even distinguish power specifications or brand models (if the training data covers these details). In practice, the model classifier can either share the same neural network as the category classification but differentiate into more granular category nodes at the final layer, or it can be an independent model. This embodiment uses a shared network structure: the intermediate features of the network are used for category discrimination, and the corresponding output layer is activated under the guidance of the category discrimination results to perform model subdivision. This design reduces redundant computation and improves efficiency.

[0060] If the classifier outputs a probability for no category that reaches the threshold (i.e., the model lacks confidence that the event belongs to any known category), then the event is considered not to belong to any load category that has appeared in the training set. Similarly, after determining the major category, if the sub-category classification results are also unclear (e.g., all models output low probabilities, or the output is "other"), then the event is considered to correspond to an unknown model. When any of the above situations occur, a rejection process is triggered: the specific device identification result is not output temporarily, but instead a "unknown load" label is generated. Simultaneously, the event's feature X and original window data are registered and saved, and a summary information is sent to the cloud to notify of the occurrence of an unknown load event. It should be noted that if the category is known but the model is unknown, the system can still output known category information as a reference while labeling it as unknown. For example, in this embodiment, when the category classification determines that the event belongs to "lighting" but the model is unknown, the result can be temporarily labeled as "lighting - unknown model," indicating that a new lighting device has appeared.

[0061] In this embodiment, a hierarchical recognition strategy effectively utilizes category priors: first, broad categories are distinguished to reduce the risk of confusion between different categories, and then models are subdivided within a smaller range, improving recognition accuracy and granularity. Furthermore, this architecture facilitates the expansion to new categories or models—when adding a new category, a new classifier output node or module is added without affecting the existing category model; when adding a new model, only the recognition capability needs to be expanded in the corresponding category subdivision module.

[0062] In another preferred embodiment, events that do not belong to the known loads are rejected using an open set approach during the hierarchical identification process. When the load identification model analyzes a load event to be identified, if the confidence scores (e.g., softmax probabilities) of all known categories output by the model are lower than a preset threshold, or if the feature vector of the event is too far from the centers of all known categories in the feature space, then the system will trigger a rejection decision. Once rejection is triggered, the system will mark the load event to be identified as an unknown load and will not output any known category or model as the identification result. This rejection mechanism based on preset threshold rules or feature space distance judgment can effectively detect and reject unknown loads in open set scenarios, avoiding misclassification.

[0063] In this embodiment, to address the issue of misidentification of unknown loads, an open-set rejection mechanism is integrated into the identification process. Whether in the category classification stage or the model segmentation stage, judgment conditions are set to detect whether the input sample belongs to the model's known distribution. Specific implementation methods include: judgment based on the confidence threshold of the classifier output—when the probability of the highest category is lower than a preset threshold, it is considered not to belong to any trained category; or judgment based on feature space distance—comparing the event feature vector with the feature centers of each known category (or model) to calculate similarity; when the nearest neighbor distance is higher than a certain threshold, the sample is determined to be an out-of-set load. For example, the model can learn a feature mean vector for each category during training; if the input features differ significantly from all categories during identification, "unknown" is output. When the identification process detects an unknown load, the rejection process is initiated: normal classification output is paused, and the event is not mistakenly classified into any existing category / model, but rather marked as unknown. Through the open-set rejection strategy, it is possible to avoid misclassifying unknown new devices as existing categories, thereby improving identification reliability and providing a mechanism for subsequent processing.

[0064] The open-set rejection mechanism includes: calculating the distance between the feature vector of the event to be identified and the feature distributions of each known category; outputting an unknown category judgment when the minimum distance is greater than a preset threshold; or detecting the maximum soft probability value output by the recognition model; when this value is lower than a confidence threshold, the sample is considered not to belong to any trained category, thus rejecting the conventional classification result. The threshold is determined through testing on a validation set or dynamically adjusted using an open-set recognition algorithm. For cases where the category is known but the model cannot be matched, the system can output the known category label in the classification result and append an "unknown model" marker, indicating only partial recognition.

[0065] like Figure 3 As shown, Figure 3 This is a schematic diagram of the hierarchical identification and rejection process in an embodiment of the present invention, including the processing sequence of load category classification and model identification. When the input event belongs to an unknown category or model, an rejection signal is output after the open set discrimination logic; if it belongs to a known category, it enters the model subdivision, and finally outputs the identification result or rejection result. The flowchart in the figure shows the logical relationship between category identification, model identification, rejection judgment and continuous learning triggering.

[0066] Furthermore, a closed loop is constructed for the continuous learning and updating of the load identification model based on the results of open-set rejection. When a load event to be identified is labeled as an unknown load, its related data is added to the training dataset. Subsequently, these event samples labeled as unknown loads can be assigned new category or model labels manually or through other auxiliary means. These samples with new labels are used, along with other historical data, to retrain or fine-tune the load identification model, thereby updating the model's parameters and enabling it to identify these previously unknown loads. Finally, this updated and more capable load identification model is redeployed to edge devices for the identification of subsequent load events. This process is repeated cyclically, continuously enhancing the model's capabilities.

[0067] For example, in the learning and update phase: Assume the initial model is deployed and can identify N known device categories / models. One day, a user adds a kitchen appliance (e.g., an air fryer) that has not appeared before. When it starts up, the system detects a power surge event ΔP of approximately 1500W. After the identification module analyzes the features, neither the category classifier nor the model classification can match (because there are no similar high-power kitchen appliance samples in the training set), so it outputs "unknown device," rejecting the identification. At this time, the gateway device sends the features and fragments of this unknown event to the cloud server. The cloud records this sample and may guess the category based on the context (e.g., inferring it as a small kitchen appliance from the power and usage time). If the system is configured with a human-machine interface, it can prompt the user to confirm the new device type, such as the user entering the device name "air fryer."

[0068] Subsequently, during daily off-peak hours (or after accumulating a certain number of unknown samples), the cloud will trigger a model update process: adding new samples to the training set and assigning them new category labels (or new model labels to existing categories). Using the expanded training data, the original model is retrained or incrementally trained. During incremental training, the model body is frozen, training only the output layer for the newly added categories, or a small learning rate is used to further train the entire model, ensuring that the new classes are learned without compromising the performance of the original classes. After training, the new model parameter version is uploaded and stored in the cloud. The gateway device will periodically check for cloud version updates, or the cloud will proactively push update notifications. After receiving the new model, the gateway replaces the old model and adds "air fryer" to the list of known devices in its local cache. The next time a kitchen appliance event of around 1500W is encountered, the model can identify the "air fryer" category / model.

[0069] Throughout the closed-loop process, users can evolve autonomously with minimal or no intervention. Notably, to prevent fluctuations caused by frequent model updates, update strategies can be configured in the cloud, such as updating only when a certain number of accumulated unknown event samples are reached or periodically (e.g., weekly). Simultaneously, data from multiple edge devices can be aggregated in the cloud for joint training (federated learning or centralized training), sharing learning results, and then synchronously distributed, enabling models across all devices to improve collectively. This provides excellent scalability and consistency for large-scale deployments (e.g., multiple households).

[0070] Therefore, a closed-loop mechanism of continuous learning enables iterative self-improvement of the model. When a rejection event occurs, the system stores the data of that event window locally and reports it to the cloud server. In the cloud, unknown events can be further processed, such as identifying the event load (determining its true device type or name) based on user feedback, and adding new samples to the training dataset. Incremental training is periodically performed on the updated dataset to update the load recognition model to include new load categories or models. Training updates can employ transfer learning or incremental learning algorithms: while retaining existing knowledge, only the model parameters corresponding to newly added categories / samples are adjusted, avoiding the forgetting of existing devices due to the introduction of new knowledge. The updated model is distributed to edge gateway devices via the network, replacing the original recognition model, enabling edge devices to recognize previously unknown loads. This forms a closed-loop cycle of "detecting the unknown – learning new knowledge – model updating." After multiple iterations, the model continuously expands the set of identifiable loads, maintaining the adaptability of the recognition system to the user's actual power consumption environment. The entire process is automatic with minimal or no user intervention, improving the model's lifecycle performance.

[0071] like Figure 4 As shown, Figure 4This is a block diagram of the continuous learning closed-loop mechanism triggered by rejection in this embodiment of the invention, including the storage and uploading of unknown event data, as well as the cloud model update and distribution process. The arrows indicate the direction of data flow and model flow: unknown events are transmitted to the cloud via the edge gateway, and the cloud updates the model and feeds it back to the edge, thus achieving a closed loop.

[0072] For the model deployment phase: The smart gateway hardware in this embodiment uses an embedded ARM processor with a main frequency of 500MHz and a current-voltage sampling analog-to-digital converter chip. Implemented in firmware written in C language, it consumes very little CPU time (<5%) during runtime and can work continuously for extended periods without missing events. The load identification module uses a compressed convolutional neural network with a model size of approximately several hundred KB. The computational cost per inference is in the tens of MFLOPS range, which can be completed in about 0.5 seconds on the ARM processor. For 10Hz sampling, an event window typically contains dozens of data points; preprocessing and feature extraction overhead are negligible, with the main time consumption occurring during model inference. The edge device runs a real-time operating system (RTOS) that prioritizes event detection and identification tasks: event detection is a high-priority real-time task, and identification is the next highest priority task. Experimental results show that when multiple events occur consecutively within a short period (e.g., a few seconds), the device can process each event window sequentially without missing events; when there are no events, the identification task does not run, thus keeping system resources in an idle, low-power state.

[0073] Regarding network connectivity, the gateway communicates with the cloud server via Wi-Fi. Only minimal data exchange occurs when an unknown event occurs or a model update is needed. For example, reporting an unknown event only requires sending a few seconds of voltage and current data (compressed to several KB) or an extracted feature vector (tens of bytes), having a minimal impact on bandwidth. Model update files are also in the hundreds of KB range, and security is ensured through TLS encryption. Edge storage stores a certain number of recent event records (e.g., the last 100 events), occupying minimal memory but providing convenience for debugging and local model updates.

[0074] like Figure 5 As shown, Figure 5 This diagram illustrates the collaborative work between edge devices and the cloud under a two-stage deployment architecture in this embodiment of the invention. The event detection module (lightweight algorithm) and load identification module (deep learning model) running on the edge device are represented by solid boxes, while the cloud training platform is represented by dashed boxes. The accompanying diagram illustrates the relationship between real-time inference performed locally at the edge, periodic training and updates performed in the cloud, and the interaction between the two through network communication.

[0075] Overall, the continuous learning and updating steps include: 1. Data Collection: The edge device temporarily stores the unknown load event data that is rejected and uploads it to the cloud periodically or in real time; the data includes the original voltage and current timing of the event window or the extracted feature vector, as well as the edge device's preliminary judgment information (such as possible category inference). 2. Sample labeling: The cloud server includes unknown events in the new sample set and uses manual feedback or auxiliary algorithms to determine the true load category or model label for the new samples; when the label cannot be clearly defined, it can be temporarily designated as a new category pending. 3. Incremental Model Training: Input the training dataset containing the newly added samples into the existing model for training and updates. Preferably, a method is adopted to keep the learned parameters of the original model stable, and only the parameters related to the newly added category / model are trained (e.g., adding new nodes to the output layer and training their connection weights, or using a multi-task learning framework to expand the model); during the training process, some samples of the original category are used at the same time to avoid forgetting. 4. Model Deployment: The updated model is distributed to the corresponding edge devices and loaded to replace the old model for operation. After a successful update, the edge devices will have the ability to recognize the new load.

[0076] The above process is repeated to continuously expand the model's knowledge base. Cloud training can be triggered immediately upon detecting new samples or performed in batches after accumulating a batch of samples. Through this closed-loop mechanism, the model can continuously learn as the usage environment evolves, ensuring long-term recognition performance.

[0077] With the aforementioned hardware and software integration, this method demonstrates stable performance in a real-world home environment. Tests show that the start-stop events of common household appliances such as air conditioners, refrigerators, and washing machines can be accurately identified, with a category recognition accuracy exceeding 95% and a model / state segmentation accuracy exceeding 85%. When two previously unknown devices (such as a new electric heater and a new kitchen appliance) are introduced, the system initially rejects them, but correctly identifies them after user confirmation of the name and model update. This proves that the present invention can reliably monitor and identify various household loads under low-frequency sampling and limited computing power conditions, and possesses excellent scalability and adaptability.

[0078] The technical process of the method is as follows: First, the event-driven model captures load changes and extracts data windows at the edge. Then, the pre-trained and optimized model extracts features and performs hierarchical identification of known loads. For unknown situations, the model is rejected and updated in the cloud. Finally, the edge model is upgraded to complete the closed loop, continuously improving the identification range and accuracy.

[0079] This method is not limited to home environments; it can also be used for monitoring electrical equipment in other low-voltage scenarios such as commercial buildings and industrial workshops. The sampling frequency and feature selection are not fixed; 10Hz is a typical value, but higher or slightly lower frequencies (such as 1Hz or 1kHz) can also be used for event-driven identification, requiring only corresponding adjustments to the model and parameters. Furthermore, in addition to voltage and current, this method is also applicable to sensor data such as power factor and harmonics, and this information can be integrated to improve performance.

[0080] The present invention also provides a low-frequency power load identification system based on event-driven and hierarchical rejection. The low-frequency power load identification system based on event-driven and hierarchical rejection provided by the present invention will be described below. The low-frequency power load identification system based on event-driven and hierarchical rejection described below can be referred to in correspondence with the low-frequency power load identification method based on event-driven and hierarchical rejection described above. Figure 6 This is a structural block diagram of the low-frequency power load identification system based on event-driven and hierarchical rejection provided by the present invention, as shown below. Figure 6 As shown, the system includes edge devices and cloud servers, specifically including: An event detection module is used to identify switching events of electrical loads in low-frequency sampled power data based on power surge detection, and to capture an identification data window containing the switching events. The training and fine-tuning module is used to build a load identification model and perform self-supervised pre-training and fine-tuning on the load identification model. The load identification module is used to perform hierarchical identification of the identification data window of the load event to be identified using the load identification model; The rejection discrimination module is used to reject events that do not belong to the known load during the hierarchical identification process by performing open set rejection. The learning and updating module is used to continuously learn and update the load identification model based on the results of the open set rejection.

[0081] The edge device includes the event detection module, the load identification module, and the rejection judgment module; the cloud server includes the training and fine-tuning module and the learning and updating module; the training and fine-tuning module and the learning and updating module are also used to distribute the load identification model, which has been pre-trained by self-supervised training or continuously learned and updated, to the edge device.

[0082] This system adopts a two-stage deployment architecture combining edge computing and cloud services to meet real-time requirements and low computational constraints, while fully utilizing cloud resources. The event detection module runs continuously with a simple and efficient algorithm (e.g., power differential detection based on sliding windows or small neural network discrimination), capable of real-time location of power transition events in a 10Hz data stream, with low computational complexity and minimal resource consumption. The load identification module (including a feature extraction network and classifier) ​​has relatively higher computational requirements, but its architecture is designed for on-demand triggering: the load identification module is only activated to process the window data when a load switch is detected. This avoids continuous computation for each sampling point, significantly reducing the average computational consumption of edge devices. The load identification module can be further optimized using model pruning, quantization, and other techniques to make it suitable for running on embedded CPUs or small NPUs, ensuring that inference latency meets real-time requirements (typically, the latency for a single event identification can be controlled within 1 second). Simultaneously, a model training and fine-tuning module and a learning and update module are deployed in the cloud to handle complex training tasks and the global model update and maintenance. The cloud can leverage its powerful computing resources to periodically train data from multiple devices (such as batch training of deep learning models and parameter tuning), and then synchronize the updated model to each edge node. The cloud and edge communicate via a secure network connection, with the edge only reporting necessary information (such as features of rejection events). Normally, it does not need to transmit all raw user data, ensuring user privacy and efficient bandwidth utilization. This two-stage architecture leverages the advantages of both the edge and the cloud: real-time response and on-site identification at the edge, and centralized training and continuous optimization in the cloud, achieving an organic combination of "local inference + cloud training".

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these 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 low-frequency power load identification method based on event-driven and hierarchical rejection, characterized in that, include: Based on power surge detection, switching events of electrical loads are identified in low-frequency sampled power data, and an identification data window containing the switching events is extracted. A load identification model is constructed, and the load identification model is subjected to self-supervised pre-training and fine-tuning. The load identification model is used to perform hierarchical identification of the identification data window of the load event to be identified; During the hierarchical identification process, events that do not belong to the known load are rejected using an open set method. Based on the results of the open set rejection, the load identification model is continuously learned and updated.

2. The low-frequency power load identification method based on event-driven and hierarchical rejection as described in claim 1, characterized in that, Based on power surge detection, switching events of electrical loads are identified in low-frequency sampled power data, and an identification data window containing these switching events is extracted, including: Voltage and current data of the power system bus are acquired using a low-frequency sampling frequency; Monitor changes in the voltage and current data, and identify switching events of electrical loads using a power surge detection algorithm; When the detected change in power or voltage exceeds a preset threshold, a switching event is determined to have occurred, and a time window centered on the moment the switching event occurs is extracted as an identification data window.

3. The low-frequency power load identification method based on event-driven and hierarchical rejection as described in claim 1, characterized in that, Before building the load identification model, the following is also included: Extract multi-dimensional features reflecting the electrical characteristics of the electrical load from the identification data window; Multiple view samples are generated by applying a physically consistent data augmentation strategy to the original feature sequence, which consists of multiple features.

4. The low-frequency power load identification method based on event-driven and hierarchical rejection as described in claim 1, characterized in that, Constructing a load identification model and performing self-supervised pre-training on the load identification model includes: Construct a feature encoding network for the load identification model; Multiple enhanced view samples of the same load event are input into the feature encoding network to obtain corresponding feature vectors; The feature encoding network is trained by a self-supervised contrastive learning algorithm, so that the feature vectors of the same load event are close to each other and the feature vectors of different load events are separated from each other. After pre-training, the feature encoding network is connected to at least one classifier to form a complete load recognition model.

5. The low-frequency power load identification method based on event-driven and hierarchical rejection according to claim 4, characterized in that, Fine-tuning the load identification model includes: Collect tagged data for known load events; The labeled known load events are input into the load identification model, the error between the classifier output and the true category label is calculated, and the model parameters are updated so that the load identification model can distinguish between predefined load categories. The output layer of the load recognition model is also configured with a mechanism for identifying unknown categories.

6. The low-frequency power load identification method based on event-driven and hierarchical rejection as described in claim 1, characterized in that, The load identification model is used to perform hierarchical identification of the identification data window of the load event to be identified, including: Input the identification data window of the load event to be identified into the load identification model, and output the load category to which the load event to be identified belongs; For each load category to which the load event to be identified belongs, the corresponding load model or device identifier is further output.

7. The low-frequency power load identification method based on event-driven and hierarchical rejection as described in claim 1, characterized in that, In the process of hierarchical identification, open-set rejection is performed on events that do not belong to the known load, including: If the confidence level of the load identification model in the load event to be identified not belonging to any known category or model is lower than the threshold, a rejection decision is triggered, the load event to be identified is marked as an unknown load, and no identification result of the existing category is output; The rejection determination is based on a preset threshold rule or feature space distance discrimination to detect and reject unknown loads in open set scenarios.

8. The low-frequency power load identification method based on event-driven and hierarchical rejection as described in claim 1, characterized in that, Based on the results of the open-set rejection, the load identification model is continuously learned and updated, including: The data of the load events to be identified, which will be marked as unknown loads, will be added to the training dataset; Event samples marked as unknown loads are assigned new categories or new model labels, and the load recognition model is retrained or fine-tuned by incorporating previously trained data to update the model parameters to recognize unknown loads. The updated load identification model is deployed to edge devices for the identification of subsequent load events to be identified.

9. A low-frequency power load identification system based on event-driven and hierarchical rejection, characterized in that, include: An event detection module is used to identify switching events of electrical loads in low-frequency sampled power data based on power surge detection, and to capture an identification data window containing the switching events. The training and fine-tuning module is used to build a load identification model and perform self-supervised pre-training and fine-tuning on the load identification model. The load identification module is used to perform hierarchical identification of the identification data window of the load event to be identified using the load identification model; The rejection discrimination module is used to reject events that do not belong to the known load during the hierarchical identification process by performing open set rejection. The learning and updating module is used to continuously learn and update the load identification model based on the results of the open set rejection.

10. The low-frequency power load identification system based on event-driven and hierarchical rejection as described in claim 9, characterized in that, The system includes edge devices and cloud servers; The edge device includes the event detection module, the load identification module, and the rejection judgment module; The cloud server includes the training fine-tuning module and the learning update module; The training fine-tuning module and the learning update module are also used to distribute the load recognition model, which has been pre-trained by self-supervised training or continuously learned and updated, to the edge device.

Images