A non-intrusive load identification method based on high-frequency power data

Abstract

The present application relates to the technical field of load identification, in particular to a non-intrusive load identification method based on high-frequency power data, which comprises the following steps: collecting high-frequency power signal data of user end voltage and current, and constructing a hierarchical data warehouse according to high-frequency power signal data processing requirements; performing signal analysis on the collected high-frequency power signal, extracting transient and steady-state characteristics, and constructing a multi-view feature matrix based on the transient and steady-state characteristics; using a multi-branch structured learning model based on the multi-view feature matrix to perform load prediction, at the same time, constructing a deep correlation analysis model to process the uncertainty problem in the high-frequency power signal data in the prediction process, and constructing a time domain correction function using the inherent periodic characteristics of the load. Through the fusion of high-frequency transient characteristics and time domain dynamics decoupling mechanism, the present application realizes high-precision and structured identification of complex composite loads and dynamic similar equipment.

Description

Technical Field

[0001] This invention relates to the field of load identification technology, and more specifically, to a non-intrusive load identification method based on high-frequency power data. Background Technology

[0002] With the continuous development of smart grids, smart homes, and industrial energy efficiency optimization technologies, accurate identification and real-time monitoring of user-side electrical equipment has become a key means to improve energy utilization efficiency and achieve refined electricity management. However, traditional invasive load monitoring methods typically require the separate installation of sensors on each electrical device, which is not only costly and complex to deploy, but also suffers from poor scalability in large-scale applications. In modern household and industrial power consumption scenarios, multiple electrical devices often operate concurrently, resulting in the collected voltage and current signals being the superposition of multiple load behaviors. Traditional methods struggle to effectively separate the individual characteristics of these superimposed signals, leading to problems such as fuzzy identification and high misjudgment rates. Furthermore, high-frequency power signals contain rich dynamic information, but due to their high data dimensionality and strong noise interference, extracting stable and interpretable physical features and constructing a structured input representation becomes a key challenge in model design. Adding to the complexity, some loads (such as air conditioner compressors and refrigerators) exhibit obvious periodic operating characteristics, which can easily be misidentified by the model as the start-up and shutdown behavior of other devices. Even if the identification result is incorrect, the confidence level of the model output may still be high, thus masking potential misjudgment risks and affecting the reliability of the system's judgment. The training of load identification models also heavily relies on high-quality labeled data. However, in practical applications, data from laboratory environments differ from real production environments, and there is a lack of a unified data organization structure and labeling system, leading to unstable model training results, limited generalization ability, and difficulty in adapting to diverse real-world deployment scenarios. Therefore, this paper proposes a non-intrusive load identification method based on high-frequency power data. Summary of the Invention

[0003] The purpose of this invention is to provide a non-intrusive load identification method based on high-frequency power data, so as to solve the problem of accurate non-intrusive load identification and monitoring in a multi-device concurrent operation environment mentioned in the background art.

[0004] To achieve the above objectives, the present invention aims to provide a non-intrusive load identification method based on high-frequency power data, comprising the following steps:

[0005] S1. Collect high-frequency power signal data of voltage and current at the user end, and build a hierarchical data warehouse according to the high-frequency power signal data processing requirements;

[0006] S2. Perform signal analysis on the acquired high-frequency power signals, extract transient and steady-state features, and construct a multi-view feature matrix based on the transient and steady-state features;

[0007] S3. Based on the multi-view feature matrix, a multi-branch structured learning model is used for load prediction. At the same time, a deep correlation analysis model is constructed to handle the uncertainty in high-frequency power signal data during the prediction process. The inherent periodic characteristics of the load are used to construct a time-domain correction function to optimize the process of the deep correlation analysis model in handling uncertainty.

[0008] S4. Based on the data and application scenarios of the hierarchical data warehouse, conduct targeted training, optimization and validation of the multi-branch structured learning model.

[0009] As a further improvement to this technical solution, in S1, the hierarchical data warehouse includes a raw data layer, a feature extraction layer, and a tag and metadata layer;

[0010] The raw data layer is used to store the acquired raw high-frequency power signal data, and the time-series database TDengine is used for storage.

[0011] The feature extraction layer stores features extracted from the raw high-frequency power signal data;

[0012] The tagging and metadata layer is used to label the collected high-frequency power signal data with equipment tags.

[0013] As a further improvement to this technical solution, in step S2, the construction of a multi-view feature matrix based on transient and steady-state features includes the following steps:

[0014] S2.1. Standardize and preprocess the format of the raw high-frequency power signal data;

[0015] S2.2 Extract transient and steady-state features characterizing the electrical behavior of the load from the original high-frequency power signal data samples;

[0016] S2.3. In the process of extracting transient and steady-state features, a multi-perspective feature extraction mechanism is introduced to capture power feature information from multiple perspectives and dimensions, and to construct an enhanced feature vector.

[0017] S2.4 Arrange all the extracted enhanced feature vectors in a fixed order to form a multidimensional feature vector, and standardize all the enhanced feature vectors.

[0018] S2.5. Combine the standardized enhanced feature vector samples by row to construct a multi-view feature matrix.

[0019] As a further improvement to this technical solution, in step S3, a multi-branch structured learning model is used for load prediction, including the following steps:

[0020] S3.1 Organize the category labels of power loads in a structured manner to construct a multi-granularity hierarchical category system;

[0021] The multi-granularity hierarchical category system includes coarse-grained categories, medium-grained categories, and fine-grained categories.

[0022] S3.2 To address the mixed high-frequency power signal phenomenon caused by the simultaneous operation of multiple loads, a multi-tag output mechanism is introduced to support structured prediction of composite loads.

[0023] S3.3. Construct a structured deep neural network model with a multi-task branch structure for different categories of hierarchical and mixed load output targets;

[0024] S3.4 Output load prediction results and structured category information.

[0025] As a further improvement to this technical solution, S3.2 introduces a multi-tag output mechanism, including the following steps:

[0026] S3.21 Define a set of payload categories as the basis for multi-label vector encoding;

[0027] S3.22. To address the aliasing problem caused by the high overlap of high-frequency load switching timings, a time-domain dynamic decoupling method is adopted to separate the power signals of multiple loads superimposed within the same time window, extract the dynamic response of each load, and generate a decoupled multi-label output vector.

[0028] S3.23. Based on the load category set and decoupling results, construct a corresponding multi-label attribution vector for each high-frequency power signal sample;

[0029] S3.24. Construct multi-label training samples using the decoupled multi-label vectors and their corresponding feature vectors.

[0030] As a further improvement to this technical solution, in S3.22, a time-domain dynamic decoupling method is used to separate power signals superimposed by multiple loads within the same time window, including the following steps:

[0031] For the detected load switch mixed events, the corresponding time window signal segment is extracted, preprocessed, and derivative time domain features are extracted from the time window.

[0032] Based on historical data and load type characteristics, a dynamic response template library covering load categories was established;

[0033] The observed hybrid high-frequency power signal is represented as a linear combination of multiple known dynamic response templates;

[0034] Construct a multi-label output vector based on the decoupling results.

[0035] As a further improvement to this technical solution, in S3.3, a structured deep neural network model with a multi-task branch structure is constructed for different categories of hierarchical and mixed load output targets, including the following steps:

[0036] S3.31. The standardized multi-view feature matrix is ​​used as the input of the deep neural network backbone model. Each enhanced feature vector sample is input into the deep neural network backbone model. Feature fusion is performed through a multilayer perceptron structure to extract shared semantic information and generate a unified high-dimensional implicit representation vector.

[0037] S3.32. Based on the multi-granularity hierarchical classification system of the load category, a parallel classification task branch is set up after the output of the deep neural network backbone model;

[0038] The classification task branches include coarse-grained category task branches, medium-grained category task branches, and fine-grained category task branches.

[0039] S3.33. Set up a multi-label branch for composite payload identification in parallel with the classification task branch, and output the probability vector of each payload;

[0040] S3.34. Design a multi-task joint loss function consisting of the cross-entropy loss of the classification task branch and the binary cross-entropy loss of the multi-label branch, to optimize structured classification and multi-label output tasks.

[0041] As a further improvement to this technical solution, step S3 involves constructing a deep correlation analysis model to address the uncertainty in high-frequency power signal data during the prediction process, including the following steps:

[0042] S3.5 Design and train temporal convolutional networks and graph neural networks to model different types of correlation information in high-frequency power signals and form diverse correlation representations;

[0043] S3.6 Construct a fusion network structure, perform dimension matching and unified encoding on the output features of temporal convolutional networks and graph neural networks, and output the corresponding confidence scores;

[0044] S3.7. Based on the data quality and model confidence of high-frequency power signals and their preprocessing characteristics, dynamically adjust the fusion weights of each deep learning model, and use the inherent periodic characteristics of the load to construct a time-domain correction function to dynamically correct the confidence score, thereby dynamically suppressing the uncertainty misjudgment caused by periodic behavior.

[0045] S3.8 Introduce a Bayesian confidence estimation mechanism at the output of the fusion network structure to predict uncertainty based on the dynamically corrected confidence score;

[0046] S3.9. Real-time acquisition of the output and dynamic weights of temporal convolutional networks and graph neural networks, while outputting multi-dimensional load prediction results and corresponding uncertainty measures.

[0047] As a further improvement to this technical solution, in step S3.7, a time-domain correction function is constructed using the inherent periodic characteristics of the load to dynamically correct the confidence score, including the following steps:

[0048] A fast Fourier transform is performed on the input multi-view feature matrix to identify the periodic components present therein, and the time-domain autocorrelation function is calculated to verify the stability and regularity of periodic fluctuations.

[0049] Based on the power variation characteristics of periodic loads during typical periods, a time-domain correction function is constructed to calculate the correction factor;

[0050] The calculated correction factor is integrated with the weight allocation mechanism in the network to correct the dynamic weights of the temporal convolutional network and the graph neural network.

[0051] As a further improvement to this technical solution, in step S4, the multi-branch structured learning model is trained, optimized, and validated based on the data and application scenarios of the hierarchical data warehouse, including the following steps:

[0052] S4.1 Extract laboratory data and production environment data from the hierarchical data warehouse to cover load signal samples of multiple categories and different operating states. Based on the labeling information of the label and metadata layer, construct a multi-label supervised training sample set and perform distribution analysis on the multi-label supervised training data samples.

[0053] S4.2. Construct a scenario-specific load combination and power demand model based on the target deployment scenario, and perform scenario-level mapping processing on the original multi-label supervised training sample set to form a structured category hierarchy adapted to the target deployment scenario;

[0054] S4.3 Using the structured deep neural network model constructed in step S3, input the standardized multi-view feature matrix, use the multi-task loss function for end-to-end training, and fine-tune the task branch structure in the structured deep neural network model;

[0055] S4.4 Introduce the latest collected production environment samples, label the new data based on pseudo-labels and manual verification, and perform periodic incremental training on the structured deep neural network model.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] 1. This non-intrusive load identification method based on high-frequency power data achieves high-precision and structured identification of complex composite loads and dynamically similar devices by integrating high-frequency transient features with a time-domain dynamic decoupling mechanism, effectively overcoming the identification bottleneck caused by the dependence on traditional steady-state features. Combined with a hierarchical data warehouse and scene-adaptive training, it ensures that the model has good generalization ability and stable performance in heterogeneous environments such as homes and industries, meeting the fine monitoring needs of multi-granularity categories and multi-label composite loads.

[0058] 2. This non-intrusive load identification method based on high-frequency power data introduces deep correlation analysis and Bayesian confidence estimation mechanisms to dynamically suppress uncertainties and misjudgments caused by external interference and periodic behavior, thereby improving the robustness and reliability of identification. Combined with automated pseudo-labels and incremental manual review and annotation, it significantly reduces operation and maintenance costs and accelerates model iteration. Relying on the efficient TDengine time series database and parallel inference, it achieves millisecond-level real-time response, meeting the stringent requirements of industrial-grade fault early warning and decision support. Attached Figure Description

[0059] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0061] Example: Please refer to Figure 1 As shown, this embodiment provides a non-intrusive load identification method based on high-frequency power data, including the following steps:

[0062] S1. Collect high-frequency power signal data of voltage and current at the user end, and build a hierarchical data warehouse according to the high-frequency power signal data processing requirements;

[0063] In this embodiment, the hierarchical data warehouse serves as the training foundation for identifying the characteristics of various types of electrical loads within the system. It can be categorized into laboratory data and production environment data based on their sources. When constructing the hierarchical data warehouse, the corresponding load types of the data are labeled, and relevant filterable information, including the amount of data, is improved to facilitate easier retrieval of corresponding data during subsequent model training. This forms a solid foundation for building a robust and stable model.

[0064] Data Source 1: Merged Production Data: Archived electricity load data that is clearly defined in the production data into the data warehouse, which is beneficial to improving the accuracy of identification of special electrical equipment.

[0065] Data Source 2: Laboratory Data: Suitable for unified learning of common electrical loads during the initial model training. This is also the main data source in training the model. The amount of data determines the accuracy of the model in subsequent identification. Furthermore, specialized training can be conducted for various complex electrical loads.

[0066] The hierarchical data warehouse includes a raw data layer, a feature extraction layer, and a tag and metadata layer.

[0067] The raw data layer is used to store the acquired raw high-frequency power signal data, including the instantaneous value sequence of voltage and current, and is stored using the time-series database TDengine.

[0068] The feature extraction layer stores features (including VI trajectory, harmonic distortion rate, and transient response) extracted from the raw high-frequency power signal data, which are used as input to the multi-branch structured learning model.

[0069] The tagging and metadata layer is used to label the collected high-frequency power signal data with equipment tags, including equipment brand, model, rated power, etc.

[0070] S2. Perform signal analysis on the acquired high-frequency power signals, extract transient and steady-state features, and construct a multi-view feature matrix based on the transient and steady-state features;

[0071] In this embodiment, a multi-view feature matrix is ​​constructed based on transient and steady-state features, including the following steps:

[0072] S2.1 Perform format standardization and preprocessing on the raw high-frequency power signal data, including: timing alignment (synchronizing voltage and current channels), filtering and denoising (using a low-pass filter), data truncation (extracting waveform segments of fixed window length around load on / off events), and data normalization to improve feature stability.

[0073] S2.2 Extract transient and steady-state features characterizing the electrical behavior of the load from the original high-frequency power signal data samples, including: time domain features (current peak value, waveform slope, transient duration), frequency domain features (harmonic energy distribution and distortion rate), as well as VI trajectory morphology indicators and instantaneous power features;

[0074] S2.3. A multi-perspective feature extraction mechanism is introduced during the extraction of transient and steady-state features to capture power characteristic information from multiple perspectives and dimensions, and construct an enhanced feature vector. The multi-perspective feature extraction mechanism specifically involves: fusing time-domain features and frequency-domain features to form a joint description of load switching response; extracting disturbance features around load switching events, including instantaneous power changes during switching and voltage response amplitude; evaluating the root mean square error and stability index of each feature during load operation to reflect the stability of load behavior; and using the VI trajectory changes of multiple sampling points in the time series for time-series fusion modeling.

[0075] S2.4 Arrange all the extracted enhanced feature vectors in a fixed order (grouped by the source type of the enhanced feature vectors) to form a multidimensional feature vector, so that each group of waveform data corresponds to a feature vector sample, where each dimension represents a specific physical feature quantity, and standardize all the enhanced feature vectors.

[0076] S2.5. Combine the standardized enhanced feature vector samples by row to construct a multi-view feature matrix. In the formula, This is a multi-view feature matrix, representing the structured input data for all samples. The total number of samples, The feature dimension (i.e., the number of features extracted) included for each sample.

[0077] S3. Based on the multi-view feature matrix, a multi-branch structured learning model is used for load prediction. At the same time, a deep correlation analysis model is constructed to handle the uncertainty in high-frequency power signal data during the prediction process. The inherent periodic characteristics of the load are used to construct a time-domain correction function to optimize the process of the deep correlation analysis model in handling uncertainty.

[0078] In this embodiment, a multi-branch structured learning model is used for load prediction, including the following steps:

[0079] S3.1 Organize the category labels of power loads in a structured manner to construct a multi-granularity hierarchical category system;

[0080] The multi-granularity hierarchical category system includes coarse-grained categories (including lighting, kitchen appliances, office equipment, and power equipment), medium-granular categories (including LED lights, microwave ovens, printers, and variable frequency air compressors), and fine-granular categories (specific brands and models, including X brand electric kettle-1500W, Y model desktop computer, etc.).

[0081] S3.2 To address the mixed high-frequency power signal phenomenon caused by the simultaneous operation of multiple loads, a multi-tag output mechanism is introduced to support structured prediction of composite loads.

[0082] The introduction of a multi-tag output mechanism includes the following steps:

[0083] S3.21 Define a set of payload categories as the basis for multi-label vector encoding;

[0084] S3.22. To address the aliasing problem caused by the high overlap of high-frequency load switching timings, a time-domain dynamic decoupling method is adopted to separate the power signals of multiple loads superimposed within the same time window, extract the dynamic response of each load, and generate a decoupled multi-label output vector.

[0085] When the switching time difference of multiple loads is less than 200ms (typical period of high-frequency sampling), their transient characteristics are highly overlapping in the time domain, resulting in: the feature vector cannot distinguish the ownership of the superimposed signal, and erroneous associations occur in multi-label vectors (mislabeling the superimposed features of microwave oven + refrigerator as a single device). Therefore, a time-domain dynamic decoupling method is designed, which utilizes the difference in dynamic response time scale of different loads (motor loads vs. resistive loads) to add dynamic behavior decoupling analysis in the feature extraction stage.

[0086] Furthermore, a time-domain dynamic decoupling method is used to separate power signals superimposed by multiple loads within the same time window, including the following steps:

[0087] For the detected load switch mixed events, the corresponding time window signal segment is extracted and preprocessed (the preprocessing operation includes filtering and noise reduction and timing alignment to ensure data quality and consistency), and derivative time domain features are extracted from the time window (derivative time domain features include current rate of change derivative, instantaneous power rate of change, and VI trajectory curvature rate of change (nonlinear behavior)).

[0088] Based on historical data and load type characteristics, a dynamic response template library covering load categories is established. Each template in the dynamic response template library represents the typical dynamic response mode of the corresponding load, including: inductive loads exhibit an exponential rise or delayed response of current, compressor loads have specific start-up and running dynamic characteristics, and the current change rate of resistive loads exhibits a rapid stabilization after a spike.

[0089] The observed hybrid high-frequency power signal is represented as a linear combination of multiple known dynamic response templates;

[0090] The linear combination is:

[0091] ;

[0092] In the formula, For the first The contribution coefficient of each load in the current hybrid high-frequency power signal (solved using the non-negative least squares method). For residual noise term, It is a mixed high-frequency power signal (i.e., the total current or total power waveform after the superposition of multiple load switch actions). For time, As a standard dynamic response template, it is a signal pattern in the form of a time function. The number of dynamic response templates, Index for load category;

[0093] Based on the decoupling results, a multi-label output vector is constructed. This vector accurately reflects the simultaneous operation of multiple loads and serves as the structured real label input for the multi-label learning model, which is used for subsequent model training and prediction.

[0094] S3.23. Based on the load category set and decoupling results, construct a corresponding multi-label attribution vector for each high-frequency power signal sample to clearly indicate all operating loads included in the sample;

[0095] S3.24. Construct multi-label training samples using the decoupled multi-label vectors and corresponding feature vectors, save synchronized timestamps and label information, support one sample to be associated with multiple load labels at the same time, realize accurate identification of composite loads, and provide multi-label training samples for multi-label branch learning in step S3.33.

[0096] S3.3. For output targets with different category levels and mixed loads, a structured deep neural network model with a multi-task branch structure is constructed. The structured deep neural network model takes the standardized multi-view power feature matrix as input. First, a unified implicit semantic representation is extracted through a shared backbone network. The backbone network is usually composed of multiple convolutional layers or fully connected layers, which are responsible for fusing and extracting multi-granularity features of the input. Based on this shared representation, the model has multiple parallel task branches: including coarse-grained classification branches, medium-grained classification branches, and fine-grained classification branches for multi-granularity levels of load categories, which output the category probability distribution of the corresponding level respectively; at the same time, there is a dedicated multi-label branch, which uses the sigmoid activation function to output multi-label probability vectors corresponding to all load categories, realizing the identification of composite loads. The overall model uses a joint loss function to weight and optimize the prediction errors of all branches, ensuring that the structured category and multi-label output tasks synergistically improve the recognition accuracy and robustness.

[0097] Specifically, for different categories and mixed load output targets, a structured deep neural network model with a multi-task branch structure is constructed, including the following steps:

[0098] S3.31. The standardized multi-view feature matrix is ​​used as input to the deep neural network backbone model (the deep neural network backbone model is a deep neural network structure used to extract unified and highly expressive latent feature representations from multi-view feature vectors. Its input is a pre-processed and standardized multi-view feature matrix. Each row represents the physical features extracted from a sample under multiple perspectives such as current, voltage, harmonics, VI trajectory, and instantaneous power. The fusion structure can use multilayer perceptron (MLP), convolutional modules, or Transformer units for feature extraction and temporal modeling. The output section is set with multiple branch channels according to application requirements, including coarse-grained load category output (resistor / inductor), fine-grained specific device type output (electric kettle / microwave oven), and multi-label prediction vectors under combined loads. In addition, uncertainty scores and confidence indices can be output in parallel). Each enhanced... Feature vector samples are input into the backbone model of a deep neural network. Feature fusion is performed through a multilayer perceptron (MLP) structure to extract shared semantic information and generate a unified high-dimensional latent representation vector. The feature fusion through the MLP structure involves dimensional alignment of the output features from multiple deep neural network sub-models (including TCN, GNN, and Attention modules) and mapping them to a unified space through linear transformations or projection layers. Subsequently, the feature vectors of all sub-models are concatenated sequentially to form a fused input vector. This vector is then input into the MLP structure, passing through several fully connected layers and non-linear activation functions (ReLU or GELU) to extract high-order combined features and capture potential interactions between models. Finally, the fused features are used in the output layer for tasks such as classification, multi-label recognition, or uncertainty estimation, resulting in structured predictions.

[0099] S3.32. Based on the multi-granularity hierarchical classification system of the load category, multiple parallel classification task branches are set up after the output of the deep neural network backbone model;

[0100] The classification task branches include coarse-grained category task branches, medium-grained category task branches, and fine-grained category task branches.

[0101] Coarse-grained categories (including device function type), medium-grained categories (including specific electrical appliance types), and fine-grained categories (including brand and model, etc.) are defined. Each task branch uses an independent fully connected layer to output the category prediction results at the corresponding granularity, forming a structured classification path.

[0102] S3.33. Set up a multi-label branch for composite payload identification in parallel with the classification task branch. Output the probability vector of each payload. This branch takes the high-dimensional latent representation vector of the backbone as input and outputs a probability vector with dimensions corresponding to the set of payload categories. Each element represents the probability of the corresponding payload category appearing in the current sample.

[0103] S3.34. Design a multi-task joint loss function composed of the cross-entropy loss of the classification task branch and the binary cross-entropy loss of the multi-label branch to optimize the structured classification and multi-label output tasks. This multi-task joint loss function consists of the cross-entropy loss of each granularity classification branch and the binary cross-entropy loss of the multi-label branch. Structured classification refers to organizing the category label system into a hierarchy (i.e., a multi-granularity hierarchical category system) and setting multiple classification task output branches according to the structured hierarchy of the payload categories to complete category prediction at different granularities. The multi-label output task refers to the model being able to predict the situation where multiple payload categories exist simultaneously in the same sample, that is, not just outputting one category, but outputting a probability vector of a set of categories (i.e., the multi-label output mechanism in step S3.2 and the parallel setting of a multi-label branch for composite payload identification in step S3.3, outputting the probability vector of each payload).

[0104] Multi-task joint loss function for:

[0105] ;

[0106] In the formula, For the cross-entropy loss of the coarse-grained classification task branch, For the cross-entropy loss of the branches in a medium-granularity classification task, For fine-grained classification task branches, the cross-entropy loss is used. For the binary cross-entropy loss of the multi-label output branch, These are the weight coefficients for the coarse-grained classification task branches. These are the weight coefficients for the branches of the medium-granularity classification task. These are the weight coefficients for the fine-grained classification task branches. The weight coefficients for the multi-label output branches;

[0107] S3.4 Output load prediction results and structured category information, including running status labels (whether there is load running), multi-granularity structured category labels (coarse-grained category, medium-grained category, fine-grained category), and multi-label output (multiple predicted candidate categories and their probabilities).

[0108] Furthermore, the deep association analysis model includes a multi-channel association modeling layer, a feature fusion layer, a dynamic weighting allocation module, and an uncertainty modeling module. The multi-channel association modeling layer contains multiple specialized sub-models (each capturing specific association information), including temporal convolutional networks and graph neural networks, outputting intermediate representation features from multiple sub-models. The feature fusion layer aligns and uniformly encodes heterogeneous features from multiple sub-models to obtain a structurally consistent joint representation, outputting a unified fused feature vector as a shared semantic representation for subsequent predictions. The dynamic weighting allocation module assigns weights to each sub-model based on input data quality (including feature completeness and confidence) and model performance, outputting a dynamic weight vector. The uncertainty modeling module estimates the uncertainty of the final prediction result, assisting in judging the credibility of the model output, and outputs classification or multi-label prediction results, confidence scores, and uncertainty measures (including prediction variance) through a Bayesian neural network.

[0109] To address the uncertainty in high-frequency power signal data during the prediction process, a deep correlation analysis model is constructed, including the following steps:

[0110] S3.5 Design and train temporal convolutional networks and graph neural networks to model different types of correlation information in high-frequency power signals and form diverse correlation representations;

[0111] The Temporal Convolutional Network (TCN) utilizes the multi-view feature matrix obtained in step S2.5 as training data, with an input format of batch size × time step × feature dimension, containing physical features of power signals under high-frequency sampling. Its architecture captures long-term dependencies through causal dilated convolutional layers and improves efficiency by combining residual connections, outputting high-dimensional temporal feature vectors and global temporal abstract features to characterize load startup and transient dynamics. The Graph Neural Network (GNN) constructs a graph structure based on multi-load interaction data within the same time window, combining device labels and metadata. Its input includes steady-state and real-time features of load nodes and edge association weights. Its architecture aggregates neighbor information through graph convolutional layers and dynamically adjusts edge weights using an attention mechanism, outputting node-level relationship-enhanced feature vectors and graph-level embedding of the overall power consumption system. The GNN effectively captures collaborative power consumption patterns among loads, enhancing the representation ability of complex combined load states and providing structured contextual information for subsequent fusion networks.

[0112] S3.6 Construct a fusion network structure, perform dimension matching and unified encoding on the output features of temporal convolutional networks and graph neural networks, use a fusion layer (splitting and combining fully connected layers) to integrate the output features of temporal convolutional networks and graph neural networks, and output the corresponding confidence scores;

[0113] S3.7. Based on the data quality and model confidence of high-frequency power signals and their preprocessing characteristics, dynamically adjust the fusion weights of each deep learning model, and use the inherent periodic characteristics of the load to construct a time-domain correction function to dynamically correct the confidence score, thereby dynamically suppressing the uncertainty misjudgment caused by periodic behavior.

[0114] In high-frequency power load identification, some periodic load devices (including air conditioner compressors, refrigerators, water heaters, etc.) repeatedly generate highly similar power fluctuation characteristics during their start-stop cycles. This can easily lead the model to misclassify them as start-stop events of other loads. More seriously, even if the model's identification result is incorrect, its output confidence level may still be high, thus masking potential prediction errors. In addition, the repetitive fluctuations of periodic loads also interfere with the perception of actual load changes, causing the model confidence level to fail to accurately reflect the reliability of the prediction. To address this, a time-domain correction function based on the inherent periodic characteristics of the load is introduced. By analyzing whether the current load behavior is within its normal periodic fluctuation range, the model's confidence level is dynamically adjusted, i.e., suppressed within the cycle and enhanced during abnormal changes, thereby improving the identification reliability and stability. This correction function integrates periodic structure information such as Fourier periodic components and autocorrelation functions, assisting the model in judging the authenticity of load behavior from a time dimension. It also serves as a correction factor in the multi-model fusion stage to dynamically adjust the model weights, preventing periodic signals from dominating the overall prediction output, effectively reducing the misclassification rate and improving the model's generalization ability in complex scenarios.

[0115] The confidence score is dynamically corrected by constructing a time-domain correction function using the inherent periodic characteristics of the load, including the following steps:

[0116] A Fast Fourier Transform (FFT) is performed on the input multi-view feature matrix to identify significant periodic components (typical start-stop cycle of an air conditioner compressor of 15–30 minutes), and the time-domain autocorrelation function is calculated. ( In the formula, Represents the first in the multi-view feature matrix The eigenvalues ​​at time t, For time indexing, The time delay (lag) variable represents the offset of the signal from itself. The degree of relevance after time, The first in the multi-view feature matrix The characteristic values ​​at each moment are used to verify the stability and regularity of periodic fluctuations;

[0117] Based on the power variation characteristics of periodic loads during typical periods, a time-domain correction function is constructed to calculate the correction factor. ( In the formula, A positive adjustment parameter, in W.-1 This is used to control the sensitivity of the correction factor; the larger the value, the more sensitive it is to changes in differences. This represents the actual power change value, indicating the currently observed power fluctuation. (This represents the typical power change predicted based on a periodic model, indicating the expected power change under normal periodic behavior).

[0118] The calculated correction factor is fused with the weight allocation mechanism in the network to correct the dynamic weights of the temporal convolutional network and the graph neural network. The corrected dynamic weights are then used. for: In the formula, Indicates the first The original dynamic weights of each network (including temporal convolutional networks and graph neural networks) reflect the current confidence or importance of the model;

[0119] S3.8. A Bayesian confidence estimation mechanism is introduced at the output of the fusion network structure. Based on dynamically corrected confidence scores, uncertainty is predicted. The uncertainty index is used to assist decision-making and identify samples with low prediction confidence. The Bayesian confidence estimation mechanism consists of a parameter uncertainty modeling module, a multiple sampling inference mechanism, and an uncertainty quantification output layer. This mechanism is based on Bayesian inference principles, treating model parameters as probability distribution variables and introducing a prior distribution to reflect the uncertainty constraints of training data on the model. During the inference phase, multiple forward propagations are performed using methods such as MC Dropout or deep ensemble to generate a prediction distribution. The mean and variance of the output are then calculated as confidence estimates of the load identification results. Finally, the mechanism outputs a structured result containing the predicted category and confidence level.

[0120] S3.9. Real-time acquisition of the output and dynamic weights of temporal convolutional networks and graph neural networks, while outputting multi-dimensional load prediction results and corresponding uncertainty measures;

[0121] In this embodiment, the multi-dimensional load prediction results include:

[0122] Multi-dimensional load prediction results (indicating whether there is load running within the time period), multi-granularity structured category labels (outputting the category labels to which the load belongs, hierarchically structured, including coarse-grained, medium-grained, and fine-grained categories; each level output is a probability distribution, and the highest probability is taken as the classification result), multi-label output results (for composite load scenarios, outputting multiple device labels and their probabilities contained in the sample, output format: a probability vector with a length equal to the total number of all device categories, each bit corresponding to the probability value of a load category), and time location information (all prediction results are bound to a specific timestamp (determined by the data collection time)).

[0123] Uncertainty measures include:

[0124] Uncertainty measure (outputting the prediction confidence score for each prediction result (including classification label and multi-label vector), dynamic adjustment result of model fusion weights (the contribution weights of different depth models (including temporal convolutional networks and graph neural networks) in the current prediction, temporal convolutional networks and graph neural networks), prediction variance or entropy (indicating the degree of uncertainty diffusion of the model under the current input; high entropy or large variance indicates that the prediction results are scattered, and thus the prediction confidence is low).

[0125] S4. Based on the data and application scenarios of the hierarchical data warehouse, conduct targeted training, optimization and validation of the multi-branch structured learning model;

[0126] In this embodiment, the multi-branch structured learning model is trained, optimized, and validated based on the data and application scenarios of the hierarchical data warehouse, including the following steps:

[0127] S4.1 Extract laboratory data and production environment data from the hierarchical data warehouse to cover load signal samples of multiple categories and different operating states. Based on the labeling information of the tag and metadata layer, construct a multi-label supervised training sample set with attributes such as brand, model, and rated power. Perform distribution analysis on the multi-label supervised training data samples to ensure that the samples are representative and balanced in dimensions such as load category, switching behavior, and superposition mode. The specific steps of the distribution analysis are as follows: First, count the frequency of each load category in the label in the sample set and calculate the label frequency distribution and its proportion in all samples. Second, analyze the label co-occurrence relationship and construct a label co-occurrence matrix to reveal the combination mode of different loads in the same time window. Third, evaluate the distribution balance of the samples in the category dimension (including load type, equipment brand, rated power, etc.), behavior dimension (including switching status, duration), and superposition mode dimension (single load / multi-load mixture).

[0128] S4.2. Based on the target deployment scenarios (including residential, commercial, and industrial scenarios), construct scenario-specific load combination and power demand models. Perform scenario-level mapping processing on the original multi-label supervised training sample set (including aggregating multiple resistive devices into a high-heat load category) to form a structured category hierarchy adapted to the target deployment scenario. The load combination and power demand model is a structured model built for a specific application scenario, mainly composed of load type sets, typical load combination configurations, combination power change characteristics, scenario labels, and time scheduling rules. This model takes historical high-frequency power data and equipment operating parameters as input, and outputs a structured list of load combination categories, power change curves corresponding to each combination, hierarchical load category mapping, and load combination activation probabilities at different time periods. This is used to characterize typical power combinations and their power demand characteristics in the target scenario. The specific construction logic is as follows: First, identify high-frequency co-occurring loads, statistically analyze frequent combination patterns, and encode them into combination configuration vectors; then, based on window statistics and feature regression, output the average power, peak power, and dynamic fluctuation characteristics of each combination per unit time; next, combine the equipment start-stop cycle to construct an operating time probability distribution, generating a combination-time joint scheduling model; finally, the model is output in a structured form.

[0129] S4.3 Using the structured deep neural network model constructed in step S3, input the standardized multi-view feature matrix, and perform end-to-end training using a multi-task loss function, which includes granular classification loss and multi-label loss, fine-tune the task branch structure in the structured deep neural network model, and combine the task weight dynamic adjustment strategy to improve the main task performance in the target scenario.

[0130] S4.4 Introduce newly collected production environment samples, label the new data using pseudo-labels and manual verification, and perform periodic incremental training on the structured deep neural network model to improve the model's adaptability to scene evolution and load changes. Deploy a validation set in the target application scenario, and quantitatively evaluate the model's load recognition accuracy, multi-label recognition accuracy, and combined load recognition accuracy. Analyze the input characteristics and predicted output of high misjudgment samples to locate model weaknesses (including periodic load and short-term concurrency).

[0131] In this embodiment, multiple algorithm models are integrated and run simultaneously in the load identification system, enabling them to work collaboratively under different scenarios or data conditions. This includes: intelligently selecting algorithms based on input data characteristics, task complexity, and computing resources to ensure the optimal model is matched for different scenarios (such as differences in sampling frequency or diverse load types); in complex or high-dimensional load environments, the system supports algorithm combinations, integrating multiple models for collaborative judgment to improve identification accuracy; simultaneously, the system drives model updates through operational feedback and on-site alarm results, achieving algorithm optimization and accuracy enhancement; in response to resource changes or task requirements, the system supports an algorithm switching mechanism, dynamically selecting the most suitable algorithm combination to adapt to the actual judgment situation; furthermore, the system should also have good algorithm scalability, enabling the rapid integration and deployment of new training models in different customer scenarios, achieving rapid replication and adaptation of algorithm capabilities, and continuously meeting the needs of changing power environments.

[0132] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A non-intrusive load identification method based on high-frequency power data, characterized in that, Includes the following steps: S1. Collect high-frequency power signal data of voltage and current at the user end, and build a hierarchical data warehouse according to the high-frequency power signal data processing requirements; S2. Perform signal analysis on the acquired high-frequency power signals, extract transient and steady-state features, and construct a multi-view feature matrix based on the transient and steady-state features; S3. Based on the multi-view feature matrix, a multi-branch structured learning model is used for load prediction. At the same time, a deep correlation analysis model is constructed to handle the uncertainty in high-frequency power signal data during the prediction process. The inherent periodic characteristics of the load are used to construct a time-domain correction function to optimize the process of the deep correlation analysis model in handling uncertainty. S4. Based on the data and application scenarios of the hierarchical data warehouse, conduct targeted training, optimization and validation of the multi-branch structured learning model; In step S3, a deep correlation analysis model is constructed to address the uncertainty in high-frequency power signal data during the prediction process, including the following steps: S3.5 Design and train temporal convolutional networks and graph neural networks to model different types of correlation information in high-frequency power signals and form diverse correlation representations; S3.6 Construct a fusion network structure, perform dimension matching and unified encoding on the output features of temporal convolutional networks and graph neural networks, and output the corresponding confidence scores; S3.

7. Based on the data quality and model confidence of high-frequency power signals and their preprocessing characteristics, dynamically adjust the fusion weights of each deep learning model, and use the inherent periodic characteristics of the load to construct a time-domain correction function to dynamically correct the confidence score, thereby dynamically suppressing the uncertainty misjudgment caused by periodic behavior. S3.8 Introduce a Bayesian confidence estimation mechanism at the output of the fusion network structure to predict uncertainty based on the dynamically corrected confidence score; S3.

9. Real-time acquisition of the output and dynamic weights of temporal convolutional networks and graph neural networks, while outputting multi-dimensional load prediction results and corresponding uncertainty measures; In step S3.7, a time-domain correction function is constructed using the inherent periodic characteristics of the load to dynamically correct the confidence score, including the following steps: A fast Fourier transform is performed on the input multi-view feature matrix to identify the periodic components present therein, and the time-domain autocorrelation function is calculated to verify the stability and regularity of periodic fluctuations. Based on the power variation characteristics of periodic loads during typical periods, a time-domain correction function is constructed to calculate the correction factor; The calculated correction factor is integrated with the weight allocation mechanism in the network to correct the dynamic weights of the temporal convolutional network and the graph neural network.

2. The non-intrusive load identification method based on high-frequency power data according to claim 1, characterized in that: In S1, the hierarchical data warehouse includes a raw data layer, a feature extraction layer, and a tag and metadata layer; The raw data layer is used to store the acquired raw high-frequency power signal data, and the time-series database TDengine is used for storage. The feature extraction layer stores features extracted from the raw high-frequency power signal data; The tagging and metadata layer is used to label the collected high-frequency power signal data with equipment tags.

3. The non-intrusive load identification method based on high-frequency power data according to claim 1, characterized in that: In step S2, a multi-view feature matrix is ​​constructed based on transient and steady-state features, including the following steps: S2.

1. Standardize and preprocess the format of the raw high-frequency power signal data; S2.2 Extract transient and steady-state features characterizing the electrical behavior of the load from the original high-frequency power signal data samples; S2.

3. In the process of extracting transient and steady-state features, a multi-perspective feature extraction mechanism is introduced to capture power feature information from multiple perspectives and dimensions, and to construct an enhanced feature vector. S2.4 Arrange all the extracted enhanced feature vectors in a fixed order to form a multidimensional feature vector, and standardize all the enhanced feature vectors. S2.

5. Combine the standardized enhanced feature vector samples by row to construct a multi-view feature matrix.

4. The non-intrusive load identification method based on high-frequency power data according to claim 1, characterized in that: In step S3, a multi-branch structured learning model is used for load prediction, including the following steps: S3.1 Organize the category labels of power loads in a structured manner to construct a multi-granularity hierarchical category system; The multi-granularity hierarchical category system includes coarse-grained categories, medium-grained categories, and fine-grained categories. S3.2 To address the mixed high-frequency power signal phenomenon caused by the simultaneous operation of multiple loads, a multi-tag output mechanism is introduced to support structured prediction of composite loads. S3.

3. Construct a structured deep neural network model with a multi-task branch structure for different categories of hierarchical and mixed load output targets; S3.4 Output load prediction results and structured category information.

5. The non-intrusive load identification method based on high-frequency power data according to claim 4, characterized in that: In S3.2, a multi-tag output mechanism is introduced, including the following steps: S3.21 Define a set of payload categories as the basis for multi-label vector encoding; S3.

22. To address the aliasing problem caused by the high overlap of high-frequency load switching timings, a time-domain dynamic decoupling method is adopted to separate the power signals of multiple loads superimposed within the same time window, extract the dynamic response of each load, and generate a decoupled multi-label output vector. S3.

23. Based on the load category set and decoupling results, construct a corresponding multi-label attribution vector for each high-frequency power signal sample; S3.

24. Construct multi-label training samples using the decoupled multi-label vectors and their corresponding feature vectors.

6. The non-intrusive load identification method based on high-frequency power data according to claim 5, characterized in that: In step S3.22, the time-domain dynamic decoupling method is used to separate power signals superimposed by multiple loads within the same time window, including the following steps: For the detected load switch mixed events, the corresponding time window signal segment is extracted, preprocessed, and derivative time domain features are extracted from the time window. Based on historical data and load type characteristics, a dynamic response template library covering load categories was established; The observed hybrid high-frequency power signal is represented as a linear combination of multiple known dynamic response templates; Construct a multi-label output vector based on the decoupling results.

7. The non-intrusive load identification method based on high-frequency power data according to claim 4, characterized in that: In step S3.3, a structured deep neural network model with a multi-task branch structure is constructed for different categories of hierarchical and mixed load output targets, including the following steps: S3.

31. The standardized multi-view feature matrix is ​​used as the input of the deep neural network backbone model. Each standardized enhanced feature vector sample is input into the deep neural network backbone model. Feature fusion is performed through a multilayer perceptron structure to extract shared semantic information and generate a unified high-dimensional implicit representation vector. S3.

32. Based on the multi-granularity hierarchical classification system of the load category, a parallel classification task branch is set up after the output of the deep neural network backbone model; The classification task branches include coarse-grained category task branches, medium-grained category task branches, and fine-grained category task branches. S3.

33. Set up a multi-label branch for composite payload identification in parallel with the classification task branch, and output the probability vector of each payload; S3.

34. Design a multi-task joint loss function consisting of the cross-entropy loss of the classification task branch and the binary cross-entropy loss of the multi-label branch, to optimize structured classification and multi-label output tasks.

8. The non-intrusive load identification method based on high-frequency power data according to claim 1, characterized in that: In step S4, the multi-branch structured learning model is trained, optimized, and validated based on the data and application scenarios of the hierarchical data warehouse, including the following steps: S4.1 Extract laboratory data and production environment data from the hierarchical data warehouse to cover load signal samples of multiple categories and different operating states. Based on the labeling information of the label and metadata layer, construct a multi-label supervised training sample set and perform distribution analysis on the multi-label supervised training data samples. S4.

2. Construct a scenario-specific load combination and power demand model based on the target deployment scenario, and perform scenario-level mapping processing on the original multi-label supervised training sample set to form a structured category hierarchy adapted to the target deployment scenario; S4.3 Using the structured deep neural network model constructed in step S3, input the standardized multi-view feature matrix, use the multi-task loss function for end-to-end training, and fine-tune the task branch structure in the structured deep neural network model; S4.4 Introduce the latest collected production environment samples, label the new data based on pseudo-labels and manual verification, and perform periodic incremental training on the structured deep neural network model.

Images