Non-intrusive load identification method for intelligent electric meter

Abstract

The invention relates to the technical field of load identification, and discloses a non-intrusive load identification method for an intelligent electric meter, and the method comprises the steps: collecting a voltage signal and a current signal of a total home-entry end of a user through the intelligent electric meter; preprocessing the collected voltage signal and current signal to obtain a preprocessed steady-state signal sequence and a transient event signal segment; and constructing and utilizing a dual-channel asynchronous feature extraction network, extracting a steady-state multi-dimensional feature vector of the load from the steady-state signal sequence, and extracting a transient-state multi-dimensional feature vector of the load from the transient-state event signal segment. According to the invention, the intelligent electric meter is used as a single signal acquisition point, complete non-intrusive load monitoring is realized, no additional sensor or monitoring equipment needs to be installed on the user side, the deployment cost and the construction complexity are greatly reduced, the privacy and safety of user power utilization are ensured, and a solid foundation is provided for large-scale popularization and application.

Description

Technical Field

[0001] This invention relates to the field of load identification technology, and in particular to a non-intrusive load identification method for smart meters. Background Technology

[0002] In the field of non-intrusive load identification, existing technologies are usually based on the analysis of the total input voltage and current signals collected by smart meters. However, traditional methods often rely on a single type of steady-state or transient features at the feature extraction level, making it difficult to comprehensively and accurately characterize the diverse and dynamically changing load characteristics. In terms of feature fusion strategies, fixed weights or simple splicing methods are often used, lacking adaptive consideration of the physical characteristics of the load, resulting in limited ability to distinguish between mixed and nonlinear loads.

[0003] Furthermore, load classification models face a trade-off between balancing the depth of time-series dependency modeling and the computational efficiency of edge computing. They may suffer from either high model complexity making deployment difficult or insufficient accuracy affecting practical application. These issues limit the accuracy, robustness, and scalability of non-intrusive load identification technology in real-world, complex power consumption scenarios.

[0004] Existing patent (CN116595488A) discloses a non-intrusive load identification method based on smart meters. Existing technologies mainly rely on the combination of directed voltage-current trajectory diagrams and fundamental wave features. The feature representation dimensions are limited and the transient characteristics of the load are not fully considered. The feature fusion method is relatively fixed and cannot adaptively adjust the weights according to the load type. The capsule neural network used has high computational complexity and is difficult to deploy efficiently on the resource-constrained meter side. Summary of the Invention

[0005] This invention provides a non-intrusive load identification method for smart meters to solve existing technical problems, addressing issues such as insufficient model deployment adaptability and the lack of a continuous optimization mechanism for the system.

[0006] To address the aforementioned technical problems, according to one aspect of the present invention, more specifically, a non-intrusive load identification method for smart meters, comprising the following steps:

[0007] S1. Collect voltage and current signals at the user's main inlet terminal through smart meters;

[0008] S2. The collected voltage and current signals are preprocessed to obtain the preprocessed steady-state signal sequence and transient event signal fragments;

[0009] S3. Construct and utilize a dual-channel asynchronous feature extraction network to extract the steady-state multidimensional feature vector of the load from the steady-state signal sequence and the transient multidimensional feature vector of the load from the transient event signal segment, respectively.

[0010] S4. Input the steady-state multidimensional feature vector and the transient multidimensional feature vector into the load characteristic sensitive dynamic fusion module, dynamically calculate and allocate fusion weights according to the preliminary load type judgment result, and generate fusion feature vector;

[0011] S5. Input the fused feature vector into the trained gated recurrent residual network for classification and recognition to obtain the final load type recognition result;

[0012] The gated cyclic residual network is constructed by combining the timing modeling capability of the gated cyclic unit with the shortcut connection structure of the residual network, and is used for efficient and accurate load classification at the meter side.

[0013] Furthermore, the preprocessing in step S2 also includes: removing the DC component and performing power frequency filtering on the acquired raw voltage and current signals, and dividing them into sequences of fixed length to form the steady-state signal sequence;

[0014] The smart meter completes the preprocessing, feature extraction, dynamic fusion, and recognition steps locally, or uploads the features to the edge gateway and the cloud to complete some of the steps.

[0015] Furthermore, the dual-channel asynchronous feature extraction network includes a steady-state feature channel and a transient feature channel;

[0016] The steady-state feature channel is used to perform at least two types of feature extraction and image transformation on the steady-state signal sequence: directed trajectory graph generation based on voltage-current trajectory, Graham angle and field image generation, and Markov transition field image generation; and convolutional encoding is performed on the generated multiple image features to obtain a steady-state multidimensional feature vector.

[0017] Furthermore, the transient feature channel is used to perform joint time-domain and frequency-domain feature extraction on transient event signal segments. The extracted features include: event duration, the ratio of peak starting current to steady-state current, the change in power factor before and after the event, and the dominant harmonic components in the event spectrum, which are then encoded to form the transient multidimensional feature vector.

[0018] Furthermore, in step S2, the detection of transient event signal segments involves calculating the sliding window differential energy E of the current signal sequence I(t). d (n) specifically refers to:

[0019]

[0020] In the above formula, L represents the length of the sliding window; when E d (n) When the threshold Th is exceeded, it is determined to be the starting point of a transient event. start And continue monitoring until Ed (n) Falling below the threshold, the event endpoint n is determined. end , cut off I(n) start :n end The voltage signal and its corresponding voltage signal are used as transient event signal segments.

[0021] Furthermore, the load-sensitive dynamic fusion module is configured to receive steady-state multidimensional feature vectors and transient multidimensional feature vectors, specifically:

[0022] S401, using a lightweight load characteristic classifier to analyze the steady-state multidimensional feature vector F s and transient multidimensional eigenvectors F t Process the data and output a probability distribution P that represents the initial type of the load.

[0023] S402. Based on the probability distribution P and a set of predefined steady-state feature importance bias parameters associated with different load types, the steady-state feature fusion weight w is dynamically calculated. s Weight w for fusion with transient features t ;

[0024] S403. The normalized steady-state feature vector and transient feature vector are weighted and summed according to their respective dynamic weights to generate a fusion feature vector F for final classification. fusion .

[0025] Furthermore, the lightweight load characteristic classifier is a fully connected neural network with a Softmax output layer, whose input is the steady-state multidimensional feature vector F. s and transient multidimensional eigenvectors F t The early splicing vectors are trained with the goal of classifying loads into K coarse-grained categories, including resistive, motor, electronic, and hybrid types.

[0026] Furthermore, the gated cyclic residual network consists of N cascaded gated cyclic residual blocks; each gated cyclic residual block performs the following operations in sequence:

[0027] S501. Temporal modeling of the input feature sequence is performed through a gated recurrent unit layer to obtain an intermediate feature sequence containing contextual information;

[0028] S502. Input the intermediate feature sequence into a residual submodule, which extracts local residual features through a one-dimensional convolutional layer and a non-linear activation function.

[0029] S503. The intermediate feature sequence is added to the local residual features through a shortcut connection parallel to the one-dimensional convolutional layer to form the output feature sequence of the gated recurrent residual block.

[0030] The gated recurrent residual network, at its end, maps the final feature sequence to a specific load type identification result through a global pooling layer and a fully connected classification layer.

[0031] Furthermore, the non-intrusive load identification method for smart meters is applied to a cloud-edge-device collaborative architecture;

[0032] Among them, the smart meter or the edge computing unit connected to it executes steps S2 to S5 to obtain the local identification result;

[0033] Simultaneously, the steady-state multidimensional feature vector, transient multidimensional feature vector, and local recognition results are uploaded to the cloud server. The cloud server is used to verify and analyze the uploaded data, generate cloud recognition results, and send error correction instructions to the meter or edge side to update the parameters of the local gated cyclic residual network or dynamic fusion module.

[0034] This invention provides a non-intrusive load identification method for smart meters. Compared with existing technologies, this method achieves the following advantages:

[0035] 1. This invention uses smart meters as a single signal acquisition point to achieve completely non-intrusive load monitoring. It eliminates the need to install any additional sensors or monitoring equipment on the user side, greatly reducing deployment costs and construction complexity. At the same time, it protects the privacy and security of users' electricity use, providing a solid foundation for large-scale promotion and application.

[0036] 2. The sliding window differential energy transient event detection algorithm adopted in this invention, combined with a dual-channel asynchronous feature extraction network, effectively enhances the ability to capture and characterize low-power loads and fast switching events, overcomes the problem of missed and false detection of events in complex power consumption scenarios by traditional methods, and improves the robustness and sensitivity of the overall identification system.

[0037] 3. The load characteristic sensitive dynamic fusion mechanism introduced in this invention can adaptively adjust the fusion weight of steady-state and transient features based on the preliminary judgment of the load type, realizing intelligent screening and enhancement at the feature level, thereby significantly improving the ability to distinguish and identify complex loads such as mixed and nonlinear loads.

[0038] 4. The gated cyclic residual network designed in this invention cleverly integrates temporal modeling and deep residual learning, achieving high-efficiency and high-precision load classification with limited computing resources. It not only fully explores the temporal dependencies of load operation, but also avoids the gradient degradation problem in deep network training, ensuring the feasibility of real-time identification at the meter side.

[0039] 5. The cloud-edge-device collaborative processing architecture constructed in this invention enables continuous iteration and performance improvement of the recognition system by completing real-time recognition at the edge and performing model verification, optimization and deployment in the cloud. At the same time, it alleviates the pressure of data transmission and cloud computing, and provides system-level support for long-term stable operation and adaptive evolution. Attached Figure Description

[0040] Figure 1 This is a flowchart of the present invention;

[0041] Figure 2 This is a simulation diagram of Embodiment 1 of the present invention;

[0042] Figure 3 This is a simulation diagram of Embodiment 2 of the present invention;

[0043] Figure 4 This is a structural diagram of Embodiment 3 of the present invention. Detailed Implementation

[0044] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, according to one aspect of the present invention, a non-intrusive load identification method for smart meters is provided, comprising the following steps:

[0046] Step 1: Collect voltage and current signals from the user's main inlet terminal using smart meters. Simultaneously collect these signals using the smart meters, utilizing their built-in high-precision sensing and analog-to-digital conversion modules to acquire raw electrical signals reflecting the load's operating status in real time. The principle behind this is relying on the smart meter as a unified data entry point, ensuring the consistency, synchronization, and integrity of the signal source, providing high-quality input for subsequent analysis. The advantage is that no additional monitoring equipment needs to be installed, achieving truly "non-intrusive" data acquisition, combining economic efficiency and deployability.

[0047] Step 2: Preprocess the acquired voltage and current signals to obtain the preprocessed steady-state signal sequence and transient event signal segments. The preprocessing also includes: removing DC components and performing power frequency filtering on the acquired raw voltage and current signals, and dividing them into fixed-length sequences to form a steady-state signal sequence. The smart meter completes the preprocessing, feature extraction, dynamic fusion, and recognition steps locally, or uploads the features to the edge gateway and the cloud to complete some steps.

[0048] The original voltage and current signals undergo DC removal, power frequency filtering, and sequence segmentation to separate the steady-state signal sequence from transient event segments. The underlying principle is based on digital signal processing technology to suppress noise and fundamental frequency interference, and the sliding window differential energy detection method to capture transient changes caused by load switching. The advantages include effectively improving signal quality, enhancing the system's sensitivity to small-amplitude transient events, and laying a time-aligned and structured data foundation for subsequent dual-channel feature extraction.

[0049] Step 3: Construct and utilize a dual-channel asynchronous feature extraction network to extract steady-state multidimensional feature vectors of the load from steady-state signal sequences and transient multidimensional feature vectors from transient event signal segments. Construct a dual-channel asynchronous feature extraction network for both steady-state and transient events. Extract steady-state multidimensional features from steady-state sequences using visualization methods, and extract transient multidimensional features from transient segments using joint time-frequency domain analysis. The underlying principle is based on the differences in the representation of load electrical characteristics at different stages, employing a heterogeneous feature extraction strategy to comprehensively capture the static and dynamic information of the load. Its advantage lies in overcoming the limitations of insufficient representation capabilities of single features, achieving complementarity and enhancement of the feature space, and improving the model's ability to identify complex loads.

[0050] Step 4: Input the steady-state and transient multidimensional feature vectors into the load-sensitive dynamic fusion module. Based on the preliminary load type judgment result, dynamically calculate and allocate fusion weights to generate a fused feature vector. A lightweight load characteristic classifier performs preliminary type judgment on the steady-state and transient features, and dynamically calculates the fusion weights based on preset category-related importance parameters to achieve weighted feature fusion. The implementation principle is based on the different degrees of dependence of load physical characteristics (such as resistive, motor, etc.) on steady-state and transient features, and adaptive weight allocation is performed. The advantage is that it avoids the suboptimal problem of fixed-weight fusion, enhances the model's feature selection ability for different load types, and improves the discriminativeness and robustness of the fused features.

[0051] Step 5: Input the fused feature vector into the trained gated recurrent residual network for classification and recognition to obtain the final load type identification result. The fused features are then input into the gated recurrent residual network, where a GRU layer captures temporal dependencies, the residual structure extracts local details, and finally, global pooling and a fully connected layer output the load type. The underlying principle combines temporal modeling and deep residual learning to achieve hierarchical feature learning and efficient classification of load operation modes. Its advantages lie in the network's ability to combine temporal awareness and feature deepening, enabling high-precision, low-latency load identification even with limited computing resources at the meter side, meeting the efficiency requirements of practical deployment.

[0052] Among them, the gated cyclic residual network is constructed by combining the timing modeling capability of the gated cyclic unit with the shortcut connection structure of the residual network, and is used for efficient and accurate load classification on the meter side.

[0053] Example 1

[0054] like Figure 2 As shown, the dual-channel asynchronous feature extraction network includes a steady-state feature channel and a transient feature channel;

[0055] The steady-state feature channel is used to perform at least two feature extraction and image transformations on steady-state signal sequences: directed trajectory graph generation based on voltage-current trajectories, Gramm's angle and field image generation, and Markov transition field image generation; and convolutional encoding is performed on the generated image features to obtain a steady-state multidimensional feature vector.

[0056] The transient feature channel is used to extract joint features in the time and frequency domains of transient event signal segments. The extracted features include: event duration, the ratio of peak starting current to steady-state current, the change in power factor before and after the event, and the dominant harmonic components in the event spectrum, which are then encoded into a transient multidimensional feature vector.

[0057] In this embodiment, for the detection of transient event signal segments, the sliding window differential energy E of the current signal sequence I(t) is calculated. d (n) specifically refers to:

[0058]

[0059] When E d (n) When the threshold Th is exceeded, it is determined to be the starting point of a transient event. start And continue monitoring until E d (n) Falling below the threshold, the event endpoint n is determined. end , cut off I(n) start :n end The voltage signal and its corresponding voltage signal are used as transient event signal segments.

[0060] The formula for the transient event detection algorithm, along with its derivation and design principles, is explained in detail below. This formula demonstrates how to robustly and sensitively capture transient current changes caused by load switching using sliding window differential energy, which is the core solution to the detection challenges of low-power loads and fast switching scenarios.

[0061] Step 1: Define the core observation – instantaneous difference

[0062] The most direct approach to transient detection is to observe the changes between adjacent sampling points of the current. The current signal sequence is defined as I(t) or in discrete form I[n], where n is the sampling point index.

[0063] First, calculate the first-order forward difference:

[0064] D[n] = I[n] - I[n-1];

[0065] In the above formula, D[n] reflects the instantaneous rate of change of current at point n. When the load changes, |D[n]| will increase significantly.

[0066] Step 2: From Differential to Differential Energy (Addressing Noise and Small Variations)

[0067] However, using only D[n] presents two problems:

[0068] 1. Sensitive to noise; individual sampling points may be affected by measurement noise, leading to false triggering.

[0069] 2. Sign cancellation: Transient processes may contain positive and negative fluctuations, which may cancel each other out if Dn| is directly accumulated.

[0070] To address the above problem, a square operation is introduced to calculate the instantaneous difference energy:

[0071] E inst [n] = |D[n]| 2 =(I[n]-I[n-1]) 2 ;

[0072] The advantages of squaring are:

[0073] Larger differences become even larger when squared, making events easier to identify. Converting all changes to positive values ​​avoids cancellation and ensures that abrupt changes in any direction contribute energy. In signal processing, the squared value is often associated with the "energy" of the signal.

[0074] Step 3: Introduce sliding window integration (to ensure event persistence)

[0075] Based solely on a single point of E inse [n] is still not stable enough. A real load switching event typically lasts for multiple sampling points (e.g., the inrush process of a motor starting). In contrast, a sharp noise pulse may only last for one or two points.

[0076] To distinguish between continuous changes and instantaneous impulses, a sliding time window of length L is introduced. The instantaneous differential energy is accumulated (integrated) within the window, and the sliding window differential energy E is defined. d (n):

[0077]

[0078] The derivation leads to the above formula:

[0079]

[0080] Among them, E d (n) represents the differential energy value calculated within a sliding window starting at point n (it quantifies the drastic change in the current signal within the window and is used to detect sudden changes caused by load switching or state changes); I(t) represents the current signal amplitude sequence collected at discrete time point t, which is one of the original inputs for load identification analysis; n represents the starting time index of the current sliding window; L represents the length of the sliding window (in units of sampling points), used to define a local interval for calculating energy changes; k represents the loop variable within the summation symbol.

[0081] Step 4: Set thresholds and detect events

[0082] For E d (n) Set a dynamic or static threshold Th. This threshold can be determined by E based on historical event-free plateau data. d (n) is determined by statistical analysis (e.g., taking the mean plus a certain number of standard deviations).

[0083] The core innovation of this embodiment lies in proposing a sliding window differential energy transient event detection algorithm and constructing a dual-channel asynchronous feature extraction network for steady-state and transient events. Its implementation principle is to accumulate the differential energy of the current sequence through a sliding window, achieving highly robust detection of load switching transient events. Furthermore, it combines the image representation of the steady-state channel with the time-frequency domain joint analysis of the transient channel to construct complementary multi-dimensional feature vectors. The advantages are a significant improvement in the detection capability for low-power loads and fast switching events, while the heterogeneous feature extraction enhances the model's representation completeness of loads with different electrical characteristics, providing richer and more stable input for subsequent identification.

[0084] Example 2

[0085] like Figure 3 As shown, the load-sensitive dynamic fusion module is configured to receive steady-state multidimensional feature vectors and transient multidimensional feature vectors, specifically:

[0086] S401, using a lightweight load characteristic classifier to analyze the steady-state multidimensional feature vector F s and transient multidimensional eigenvectors F t The process is performed to output a probability distribution P = [p1, p2, ..., p] that represents the initial type of load. k ];

[0087] Where P = [p1, p2, ..., p k [] represents the initial type probability distribution vector output by the lightweight classifier. K is the total number of coarse-grained load classes; p kThis represents the probability that the input sample belongs to the k-th class.

[0088] S402. Based on the probability distribution P and a set of predefined steady-state feature importance bias parameters associated with different load types, the steady-state feature fusion weight w is dynamically calculated. s Weight w for fusion with transient features t The calculation formula is:

[0089]

[0090] Where, α k σ represents a learnable steady-state feature importance bias parameter associated with the k-th type of load (it encodes prior knowledge, such as that the steady-state features of certain types of loads (e.g., resistive loads) are more discriminative); σ is the Sigmoid activation function, used to compress the weighted summation value into the (0, 1) interval to ensure the rationality of the fusion weights; w s This represents the calculated steady-state eigenvector F. s Dynamic fusion weights; w t This represents the calculated transient eigenvector F. t Dynamic fusion weights.

[0091] S403. The normalized steady-state feature vector and transient feature vector are weighted and summed according to their respective dynamic weights to generate a fusion feature vector F for final classification. fusion Its formula is:

[0092] F fusion =w s ·Norm(F s )+w t ·Norm(F t );

[0093] Where Norm(·) represents the layer normalization operation, used to normalize the feature vector F s and F t Standardization is applied to achieve zero mean and unit variance to improve training stability and fusion performance; F fusion This indicates that the final fused feature vector generated after dynamic weighted summation will be used as the input to the subsequent accurate classification network.

[0094] In this embodiment, the lightweight load characteristic classifier is a fully connected neural network with a Softmax output layer, and its input is the steady-state multidimensional feature vector F. s and transient multidimensional eigenvectors F t The early splicing vectors are trained with the goal of classifying loads into K coarse-grained categories, including resistive, motor, electronic, and hybrid types.

[0095] The design of the dynamic weighting formula stems from an in-depth analysis and innovative solution to a key technical problem in non-invasive load identification. Its derivation follows a logical chain from the physical problem to mathematical modeling, as follows:

[0096] Step 1: Obtain the initial type preference of the load

[0097] First, a quick, lightweight judgment is needed to determine what coarse-grained type the current load might be. This is the basis for dynamic adjustment.

[0098] Input: The extracted steady-state feature vector F s and transient eigenvector F t .

[0099] Operation: Concatenate them early and input them into a lightweight load characteristic classifier (e.g., a shallow fully connected network);

[0100] Output: A preliminary probability distribution of the type P = [p1, p2, ..., p k ], where K is the preset number of coarse-grained categories (e.g., resistive, motor, electronic, etc.), p k This represents the probability that the current load belongs to the k-th class, and

[0101] Step 2: Define the feature importance parameters for type dependencies.

[0102] A way is needed to encode the prior or learned knowledge of "the degree to which different load types depend on steady-state characteristics".

[0103] Therefore, a learnable parameter α is defined for each coarse-grained type k. k This is called the "steady-state characteristic importance bias parameter".

[0104] Where, α k The physical meaning of is: the cardinality of the inherent importance of the steady-state characteristics of the k-th type of load relative to the transient characteristics, learned from the model training data;

[0105] If α k A large positive number indicates that the identification of this type of load mainly relies on steady-state characteristics (such as resistive loads);

[0106] If α k A small or negative number indicates that the identification of this type of load mainly relies on transient characteristics, while steady-state characteristics contribute little or may even cause interference (such as motor loads).

[0107] These α kThe parameters are learned through backpropagation during the model training phase and are part of the model's knowledge.

[0108] Step 3: Dynamically calculate steady-state characteristic weights

[0109] Given the current load type tendency P and the steady-state characteristic importance base α for each load type, a weighted average of these base values ​​is taken based on the probability that the current load belongs to each category, thus obtaining a comprehensive steady-state characteristic importance assessment value for the current specific load.

[0110] Then, calculate the weighted sum:

[0111] This operation means that if the model is very confident that the current load is type k (i.e., p) k If z ≈ 1), then z ≈ α k The default importance of this type of load is directly adopted.

[0112] If the model's judgment is ambiguous (multiple p) k If all of them are not small, then z is the expected value of all possible types, which is a smooth and stable strategy.

[0113] To make this importance assessment a valid weight within the range (0, 1), the sigmoid function σ(x) = 1 / (1+e) is used. -x Perform mapping.

[0114] Calculate the steady-state characteristic weights:

[0115] The Sigmoid function maps real numbers z to the interval (0, 1), perfectly matching the weight definition. It provides non-linearity, meaning that when z is positive and large, w... s Approaching 1 (emphasizing steady-state characteristics); when z is negative and its absolute value is large, w s Approaching 0 (weakening steady-state characteristics); near z=0, the weight changes are sensitive.

[0116] Step 4: Determine the transient feature weights

[0117] To maintain a constant total weight to avoid scaling changes and to reflect the complementarity of steady-state and transient features, the weights of transient features are defined as the complements of the steady-state weights:

[0118] w t =1-w s ;

[0119] This means the system automatically assigns the remaining importance to transient features. When steady-state features are emphasized (w s (Increase), and transient characteristics are naturally weakened (w) t(Decrease), and vice versa.

[0120] Step 5: Generate fused feature vectors

[0121] The normalized features are weighted and fused using dynamically calculated weights, specifically as follows:

[0122] F fusion =w s ·Norm(F s )+w t ·Norm(F t );

[0123] The layer normalization Norm(·) in the above formula is used to eliminate F s and F t The potential differences in dimensions and scales make weighted summation more equitable and efficient.

[0124] The innovation of this embodiment lies in its load-characteristic-sensitive dynamic fusion mechanism, which achieves adaptive weight allocation through a lightweight load characteristic classifier and a type-related steady-state feature importance bias parameter. The principle is to first perform a coarse-grained type determination of the load, and then dynamically calculate and weight-fuse the features of both based on the different degrees of dependence of different load types on steady-state and transient features. The advantage lies in breaking the limitations of traditional fixed-weight fusion, realizing intelligent feature selection based on load characteristics, thereby effectively improving the robustness and discriminative power of recognition in complex load scenarios, especially demonstrating stronger adaptability in the recognition of mixed and nonlinear loads.

[0125] Example 3

[0126] like Figure 4 As shown, the gated cyclic residual network consists of N cascaded gated cyclic residual blocks; each gated cyclic residual block performs the following operations in sequence:

[0127] S501. Temporal modeling of the input feature sequence is performed through a gated recurrent unit layer to obtain an intermediate feature sequence containing contextual information; the implementation principle of this step is as follows:

[0128] A gated recurrent unit (GRU) is a variant of a recurrent neural network used to process sequential data. It effectively captures long-short dependencies in time series by introducing update and reset gates to control the retention and forgetting of historical information. In load identification scenarios, the current and voltage signals generated during appliance switching and operation exhibit significant temporal characteristics. The GRU can perform context-aware encoding on these signal sequences, extracting temporal patterns related to the load type.

[0129] The input to this step is a feature sequence, typically a multi-dimensional feature vector passed from the previous module, arranged by time step. The GRU layer processes the features at each time step sequentially, and its internal operation includes:

[0130] 1) Based on the current input and the hidden state of the previous time step, determine how much past information needs to be forgotten. This helps the model quickly adapt to new signal patterns when load conditions change abruptly (such as appliance startup).

[0131] 2) Control how much historical information should be retained and how much new information should be absorbed in the current hidden state. This mechanism enables the model to maintain a stable memory of the load type during steady-state operation, while updating the state in a timely manner when transient events occur.

[0132] 3) Combine the result of the reset gate with the current input to generate a temporary hidden representation for subsequent state updates.

[0133] 4) By updating the gate, the previous hidden state and the candidate hidden states are weighted and fused to output the hidden state at the current time step. This state contains the context information up to the current time.

[0134] Finally, the GRU layer outputs a hidden state sequence of the same length as the input sequence, called the "intermediate feature sequence". This sequence not only retains the original feature information, but also incorporates the time-dimensional dependencies.

[0135] S502. Input the intermediate feature sequence into a residual submodule, which extracts local residual features through a one-dimensional convolutional layer and a non-linear activation function; the implementation principle of this step is as follows:

[0136] After obtaining intermediate feature sequences with contextual information, this step introduces a residual submodule to further extract more local and subtle feature patterns (such as harmonic features of specific loads, instantaneous power fluctuations, etc.) from the sequence. This module adopts a one-dimensional convolutional structure, specifically designed to capture the local correlations of the sequence within a short time window. The core idea of ​​residual learning is to enhance the model's expressive power and alleviate the gradient vanishing problem in deep networks by learning the residuals (differences) between the input and output.

[0137] The intermediate feature sequence is input into the residual submodule, and its processing flow is as follows:

[0138] 1) Perform sliding convolution operations on the input sequence using one or more one-dimensional convolution kernels. Each convolution kernel is responsible for detecting a specific local feature pattern (such as current rising edge, periodic fluctuation, etc.), and generates a series of feature maps through the convolution operation.

[0139] 2) The convolution output is processed by a non-linear activation function (such as ReLU), which introduces the non-linear expressive power of the model, enabling it to fit more complex load feature functions.

[0140] 3) As needed, features may be further normalized and regularized using techniques such as batch normalization and dropout layers to enhance the model's generalization ability and stability.

[0141] The output of the residual submodule is called "local residual feature", which highlights the local details in the intermediate feature sequence that are discriminative for load classification.

[0142] S503. The intermediate feature sequence is added to the local residual features through a shortcut connection parallel to the one-dimensional convolutional layer to form the output feature sequence of the gated recurrent residual block.

[0143] In this process, the gated recurrent residual network's final end uses a global pooling layer and a fully connected classification layer to map the final feature sequence into a specific load type identification result.

[0144] To further facilitate information flow and avoid feature degradation in deep networks, this step employs a shortcut connection structure from residual networks. This connection directly adds the output of step S501 (intermediate feature sequence) to the output of step S502 (local residual features), achieving direct feature fusion. This design ensures that even if the changes learned by the residual submodules are small, the original information can be losslessly passed to subsequent layers, which is beneficial for gradient backpropagation and stable model training. Specifically, the following operations are performed:

[0145] 1) Ensure that the intermediate feature sequence and the local residual features are completely consistent in terms of time steps and feature dimensions. Dimension matching is usually achieved by adjusting the convolution stride, padding method, or using linear projection.

[0146] 2) Perform a time-step, feature-dimension-by-feature addition operation on both. This operation essentially superimposes the local feature adjustment learned by the residual submodule onto the original context features to form an enhanced feature representation.

[0147] 3) The sum is used as the final output feature sequence of the current gated cyclic residual block and passed to the next block or subsequent processing layer.

[0148] Through this cascaded mechanism of "temporal modeling → local feature extraction → feature enhancement", each gated recurrent residual block can progressively extract more discriminative load feature representations. The stacking of multiple such blocks constitutes a deep gated recurrent residual network, which ultimately maps sequence features to specific load type labels through global pooling and fully connected layers, achieving high-precision, high-efficiency, and non-intrusive load identification.

[0149] In this embodiment, a non-intrusive load identification method for smart meters is applied to a cloud-edge-device collaborative architecture;

[0150] In this process, the smart meter or the edge computing unit connected to it executes steps S2 to S5 to obtain the local recognition result; at the same time, the steady-state multidimensional feature vector, the transient multidimensional feature vector and the local recognition result are uploaded to the cloud server; the cloud server is used to verify and analyze the uploaded data, generate the cloud recognition result, and send the error correction instruction to the meter or edge side to update the parameters of the local gated cyclic residual network or dynamic fusion module.

[0151] The innovation of this embodiment lies in the design of a gated recurrent residual network, which deeply integrates the temporal modeling capability of the gated recurrent unit (GRU) with the shortcut connection structure of the residual network. Its implementation principle is to capture the long-term temporal dependencies of load operation through the GRU layer, then extract local detailed features through a one-dimensional convolutional residual submodule, and finally achieve feature enhancement and smooth gradient propagation through shortcut connections. The advantage is that the network combines temporal awareness with the efficiency of deep feature extraction. It can effectively model the dynamic changes during load switching and operation, while avoiding the gradient vanishing and feature degradation problems found in deep networks. It can still achieve high-precision and high-efficiency load classification even in the context of limited resources at the meter side.

[0152] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A non-intrusive load identification method for smart meters, characterized in that, Includes the following steps: S1. Collect voltage and current signals at the user's main inlet terminal through smart meters; S2. The collected voltage and current signals are preprocessed to obtain the preprocessed steady-state signal sequence and transient event signal fragments; S3. Construct and utilize a dual-channel asynchronous feature extraction network to extract the steady-state multidimensional feature vector of the load from the steady-state signal sequence and the transient multidimensional feature vector of the load from the transient event signal segment, respectively. S4. Input the steady-state multidimensional feature vector and the transient multidimensional feature vector into the load characteristic sensitive dynamic fusion module, dynamically calculate and allocate fusion weights according to the preliminary load type judgment result, and generate fusion feature vector; S5. Input the fused feature vector into the trained gated recurrent residual network for classification and recognition to obtain the final load type recognition result; The gated cyclic residual network is constructed by combining the timing modeling capability of the gated cyclic unit with the shortcut connection structure of the residual network, and is used for efficient and accurate load classification at the meter side.

2. The non-intrusive load identification method for smart meters according to claim 1, characterized in that: The preprocessing in step S2 further includes: removing DC components and performing power frequency filtering on the acquired raw voltage and current signals, and dividing them into sequences of fixed length to form the steady-state signal sequence; the smart meter completes the preprocessing, feature extraction, dynamic fusion and recognition steps locally.

3. The non-intrusive load identification method for smart meters according to claim 1, characterized in that: The dual-channel asynchronous feature extraction network includes a steady-state feature channel and a transient feature channel; The steady-state feature channel is used to perform at least two types of feature extraction and image transformation on the steady-state signal sequence: directed trajectory graph generation based on voltage-current trajectory, Graham angle and field image generation, and Markov transition field image generation; and convolutional encoding is performed on the generated multiple image features to obtain a steady-state multidimensional feature vector.

4. The non-intrusive load identification method for smart meters according to claim 3, characterized in that: The transient feature channel is used to extract joint features in the time and frequency domains of transient event signal segments. The extracted features include: event duration, the ratio of peak starting current to steady-state current, the change in power factor before and after the event, and the dominant harmonic components in the event spectrum, which are then encoded to form the transient multidimensional feature vector.

5. The non-intrusive load identification method for smart meters according to claim 4, characterized in that: In step S2, the detection of transient event signal segments involves calculating the sliding window differential energy E of the current signal sequence I(t). d (n) specifically refers to: In the above formula, L represents the length of the sliding window; when E d (n) When the threshold Th is exceeded, it is determined to be the starting point of a transient event. start And continue monitoring until E d (n) Falling below the threshold, the event endpoint n is determined. end , cut off I(n) start :n end The voltage signal and its corresponding voltage signal are used as transient event signal segments.

6. The non-intrusive load identification method for smart meters according to claim 1, characterized in that: The load-sensitive dynamic fusion module is configured to receive steady-state multidimensional feature vectors and transient multidimensional feature vectors, specifically: S401, using a lightweight load characteristic classifier to analyze the steady-state multidimensional feature vector F s and transient multidimensional eigenvectors F t Process the data and output a probability distribution P that represents the initial type of the load. S402. Based on the probability distribution P and a set of predefined steady-state feature importance bias parameters associated with different load types, the steady-state feature fusion weight w is dynamically calculated. s Weight w for fusion with transient features t ; S403. The normalized steady-state feature vector and transient feature vector are weighted and summed according to their respective dynamic weights to generate a fusion feature vector F for final classification. fusion .

7. The non-intrusive load identification method for smart meters according to claim 6, characterized in that: The lightweight load characteristic classifier is a fully connected neural network with a softmax output layer, whose input is the steady-state multidimensional feature vector F. s and transient multidimensional eigenvectors F t The early splicing vectors are trained with the goal of classifying loads into K coarse-grained categories, including resistive, motor, electronic, and hybrid types.

8. The non-intrusive load identification method for smart meters according to claim 1, characterized in that: The gated cyclic residual network consists of N cascaded gated cyclic residual blocks; each gated cyclic residual block performs the following operations in sequence: S501. Temporal modeling of the input feature sequence is performed through a gated recurrent unit layer to obtain an intermediate feature sequence containing contextual information; S502. Input the intermediate feature sequence into a residual submodule, which extracts local residual features through a one-dimensional convolutional layer and a non-linear activation function. S503. The intermediate feature sequence is added to the local residual features through a shortcut connection parallel to the one-dimensional convolutional layer to form the output feature sequence of the gated recurrent residual block. The gated recurrent residual network ends with a global pooling layer and a fully connected classification layer, which maps the final feature sequence to a specific load type identification result.

9. The non-intrusive load identification method for smart meters according to claim 1, characterized in that: The non-intrusive load identification method for smart meters is applied to a cloud-edge-device collaborative architecture. Among them, the smart meter or the edge computing unit connected to it executes steps S2 to S5 to obtain the local identification result; Simultaneously, the steady-state multidimensional feature vector, transient multidimensional feature vector, and local recognition results are uploaded to the cloud server; the cloud server is used to verify and analyze the uploaded data and generate cloud recognition results.

Images