Non-intrusive power load identification method

Abstract

The invention discloses a non-intrusive power load identification method, which relates to the technical field of power load identification, and comprises the following steps of: removing background noise by adopting adaptive filtering, separating a target load signal from an interference signal by combining a modal decomposition method, ensuring the purity of an input signal, performing normalization processing on a voltage signal and a current signal at the same time, and obtaining a non-intrusive power load identification result; the influence of power fluctuation during operation of different devices is suppressed, and stable input data is provided for steady-state and transient-state feature extraction; a maximum correlation minimum redundancy mRMR method is introduced, mutual information between features and load categories is calculated, redundancy calculation between the features is combined, optimal features are selected, interference of noise on classification is reduced, and meanwhile calculation efficiency is improved; in the classification stage, a convolutional neural network CNN is combined with a long-short term memory network LSTM, the CNN is responsible for extracting a spatial pattern of load features, and the LSTM processes time sequence dependence of load signals, so that the perception capability of the model for load state changes is improved.

Description

Technical Field

[0001] The invention relates to the technical field of electric load identification, and in particular to a non-intrusive electric load identification method. Background Art

[0002] In the non-invasive power load identification (NILM) process, it is extremely important whether the load characteristics extracted from the total current signal are accurate. In practical applications, affected by the noise interference of the power system, the traditional identification method has poor adaptability in a low signal-to-noise ratio environment and the recognition accuracy decreases.

[0003] Existing methods mainly optimize feature extraction, improve classification models or introduce deep learning algorithms, but they basically assume that the input signal quality is high and ignore the interference problem in the signal acquisition process. In complex environments such as industrial scenarios and commercial buildings, voltage and current signals are easily interfered by non-target loads, such as large equipment harmonics, transient current fluctuations or grid noise. Such interference leads to load feature aliasing, making it difficult for classifiers to accurately distinguish different loads, greatly reducing the recognition accuracy.

[0004] In order to reduce noise interference, some methods adopt filtering or feature fusion strategies, such as wavelet transform, to separate signals of different frequency bands to weaken high-frequency noise, or use principal component analysis PCA or maximum correlation minimum redundancy (mRMR) to optimize feature selection; however, filtering methods are likely to cause loss of useful signals and affect feature integrity, while feature fusion strategies have poor adaptability in different application scenarios and are difficult to maintain stable effects; it can be seen that in the face of a complex power environment, how to effectively weaken the impact of noise while ensuring that key information is not lost is still a difficult problem that needs to be solved urgently by the current NILM method. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a non-intrusive power load identification method to solve the problem that the existing method relies on signal quality and features aliasing under noise interference, resulting in reduced load identification accuracy and affecting classification reliability.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: An embodiment of the present invention provides a non-intrusive power load identification method, which includes: Step S1, collecting voltage signals and current signals, obtaining corresponding waveform data, and performing normalization and removing DC offset; Step S2, extracting load characteristics based on the voltage signal and current signal preprocessed in step S1; Step S3, for load characteristics, the maximum relevance minimum redundancy method is used to screen the characteristic data; Step S4, based on the feature data processed in step S3, inputting into a classification model, using a convolutional neural network (CNN) combined with a long short-term memory (LSTM) network to perform load classification; In step S4, CNN is used to extract spatial features and LSTM is used to extract temporal features; In step S4, the model parameters are adjusted in combination with adversarial training to reduce the impact of signal interference on the classification results; Step S5, calculating the classification confidence according to the load category in step S4, performing secondary discrimination on low-confidence samples, and adjusting the classification result in combination with the dynamic matching degree; Step S6, based on the adjusted load category classification result, establish a load status record, perform operation mode analysis, and monitor abnormal load status.

[0008] As a preferred solution of the non-intrusive power load identification method of the present invention, wherein: In step S1, adaptive filtering is used to suppress noise on the current signal to remove background noise; In step S1, the voltage signal and the current signal are decomposed by using a modal decomposition method to extract the target load signal and suppress the interference signal.

[0009] As a preferred solution of the non-intrusive power load identification method of the present invention, in step S1, normalization and DC offset removal are performed, specifically: Collect voltage and current signals and perform preprocessing. Calculate the average value of the acquired signal and subtract it from the original signal. Calculate the maximum amplitude of the signal and divide all data points of the signal by the maximum amplitude to normalize the signal amplitude range to the standardized interval.

[0010] As a preferred solution of the non-intrusive power load identification method of the present invention, wherein: the load characteristics include steady-state characteristics and transient characteristics, Steady-state characteristics include power factor, active power, reactive power and harmonic characteristics. Transient characteristics include power mutation point, instantaneous power change rate and dynamic power factor.

[0011] As a preferred solution of the non-intrusive power load identification method of the present invention, the step of extracting load characteristics is as follows: Calculate the steady-state characteristics of the load, including power factor, active power, reactive power and harmonic characteristics. The power factor reflects the phase relationship between current and voltage, the active power represents the energy actually consumed by the load, the reactive power measures the ineffective energy exchange in the power grid, and the harmonic characteristics are used to measure the degree of harmonic distortion of the load. Identify the transient characteristics of the load, including power mutation point, instantaneous power change rate and dynamic power factor. The power mutation point is used to identify the start and stop status of the load. The instantaneous power change rate represents the dynamic characteristics of the load. The dynamic power factor reflects the change of the power factor of the load at different time points.

[0012] As a preferred solution of the non-intrusive power load identification method described in the present invention, in which: in step S3, the filtered feature data is weighted calculated in combination with the adaptive feature enhancement method, and the feature data in a low signal-to-noise ratio environment is weighted processed using the attention mechanism.

[0013] As a preferred solution of the non-intrusive power load identification method of the present invention, the step of screening the characteristic data of the load characteristics by using the maximum correlation minimum redundancy method is as follows: For each load characteristic , calculate its correlation with the category label The mutual information of is calculated as: , in, Indicates Load characteristics, Indicates the load category, Representation characteristics With category The mutual information between Features Possible values, For Category Possible values, Features The value is And the category is The joint probability of and Respectively represent characteristics and categories The marginal probability of For all feature pairs , calculate mutual information, the formula is: , in, Indicates Load characteristics, Representation characteristics and The mutual information between Features Possible values, Features The value is And features The value is The joint probability of Calculate features Overall rating: , in, Representation characteristics The overall rating of is the selected feature set, is the number of selected features, according to Sort in descending order and select the first Features are used as the final screening results; The filtered features are weighted, and the weights By feature importance index Normalized, the normalized formula is: , Finally, the weighted features are calculated: , in, It is the final filtered and weighted feature set.

[0014] As a preferred solution of the non-intrusive power load identification method described in the present invention, the step of inputting the classification model based on the feature data processed in step S3 and using the convolutional neural network CNN combined with the long short-term memory network LSTM to classify the load is as follows: Constructing feature matrix As input to the CNN-LSTM model: , in, Represents the input feature matrix, containing The feature sequence of time steps, Represents the time step The load characteristic vector at Use a one-dimensional convolution kernel to perform convolution operation on the input features: , in, is the feature map after convolution, The convolution kernels are located at The output at For the The convolution kernel weights, is the input feature matrix at position The value at For the The bias term of the convolution kernel, is the size of the convolution kernel, is the activation function, To perform pooling operation: , in, is the feature vector after pooling, and the maximum pooling operation is used to extract information. LSTM calculates the current time step The hidden state of: , in, is the time step The hidden state of is the activation function of the LSTM unit, and are the weight matrices for input and state, respectively, is the hidden state of the previous time step, is the bias term of the LSTM unit, Use the adversarial sample to perturb the input and calculate the perturbed input. The formula is: , in, For adversarial samples, is the disturbance intensity coefficient, is the loss function, Represents the loss function For input The gradient of represents the symbolic function, Finally, CNN-LSTM generates load classification results .

[0015] As a preferred solution of the non-intrusive power load identification method described in the present invention, the steps of calculating the classification confidence according to the load category in step S4, performing secondary discrimination on low-confidence samples, and adjusting the classification results in combination with the dynamic matching degree are as follows: Calculate the classification confidence, and the classification model outputs the category probability vector : , in, is the classification probability vector, containing The predicted probability of each category, Indicates The predicted probability of the class, Calculating classification confidence : , in, is the confidence level, For low confidence samples, calculate their similarity with the nearest neighbor samples : , in, The current sample and The similarity of neighbor samples is The nearest neighbor sample The eigenvector at If the similarity If it is less than the set threshold, the classification label is adjusted; Calculate the category match: , in, For Category The matching degree, if If it is lower than the set threshold, the classification result will be corrected.

[0016] As a preferred solution of the non-intrusive power load identification method of the present invention, the steps of establishing a load status record based on the adjusted load category classification result, performing operation mode analysis, and monitoring abnormal load status are as follows: According to the load classification result adjusted in step S5, the operating status of the load is recorded and a time series data storage is established. Perform pattern analysis on the recorded data to identify the typical operating modes of the load, including steady-state operation, intermittent operation and sudden changes. Statistical methods and machine learning models are used to detect abnormal load conditions, including power mutations and operating time abnormalities, and historical data is combined to classify abnormalities. The detected abnormal load is analyzed in combination with expert rules, and abnormal alarms are triggered or classification strategies are adjusted.

[0017] The beneficial effects of the present invention are as follows: the present invention adopts adaptive filtering to remove background noise, and combines the modal decomposition method to separate the target load signal and the interference signal, thereby ensuring the purity of the input signal, and at the same time normalizing the voltage and current signals to suppress the influence of power fluctuations during the operation of different equipment, and providing stable input data for steady-state and transient feature extraction; introducing the maximum relevant minimum redundancy (mRMR) method to calculate the mutual information between features and load categories, and combining the redundancy calculation between features to select the optimal features, thereby reducing the interference of noise on classification and improving the calculation efficiency.

[0018] The present invention adopts a convolutional neural network (CNN) combined with a long short-term memory (LSTM) network in the classification stage. CNN is responsible for extracting the spatial pattern of load characteristics, and LSTM processes the temporal dependency of load signals, thereby improving the model's perception of load state changes. On this basis, adversarial training is introduced to enhance the model's adaptability to abnormal data in a low signal-to-noise ratio environment by constructing perturbation samples, thereby improving the robustness of classification. In order to further reduce misclassification, the classification confidence is calculated after classification, and a secondary discrimination is performed on low-confidence samples. The classification results are adjusted by the nearest neighbor sample similarity and dynamic matching, thereby reducing the impact of noise on classification decisions.

[0019] The present invention establishes a load status record and identifies the load operation mode in combination with time series analysis. At the same time, an abnormality detection mechanism is introduced to identify abnormal load states such as power mutation and abnormal operation time, and classifies them in combination with historical data.

[0020] In summary, the present invention significantly improves the robustness, classification accuracy and anomaly detection capability of non-intrusive power load identification in a low signal-to-noise ratio environment through signal denoising, feature screening optimization, deep learning classification enhancement and post-classification adjustment strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0022] Figure 1 It is a flow chart of the non-intrusive power load identification method of the present invention. DETAILED DESCRIPTION

[0023] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0024] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0026] Example 1, reference Figure 1 , this embodiment provides a non-intrusive power load identification method, comprising the following steps: Step S1, collecting voltage signals and current signals, obtaining corresponding waveform data, and performing normalization and removing DC offset; In step S1, adaptive filtering is used to suppress noise on the current signal to remove background noise; In step S1, the voltage signal and the current signal are decomposed by using a modal decomposition method to extract the target load signal and suppress the interference signal; In step S1, normalization and DC offset removal are performed, specifically: Collect voltage and current signals and perform preprocessing. Calculate the average value of the acquired signal and subtract it from the original signal. Calculate the maximum amplitude of the signal and divide all data points of the signal by the maximum amplitude to normalize the signal amplitude range to a standardized interval; Step S2, extracting load characteristics based on the voltage signal and current signal preprocessed in step S1; Load characteristics include steady-state characteristics and transient characteristics. Steady-state characteristics include power factor, active power, reactive power and harmonic characteristics. Transient characteristics include power mutation point, instantaneous power change rate and dynamic power factor; The steps to extract load characteristics are: Calculate the steady-state characteristics of the load, including power factor, active power, reactive power and harmonic characteristics. The power factor reflects the phase relationship between current and voltage, the active power represents the energy actually consumed by the load, the reactive power measures the ineffective energy exchange in the power grid, and the harmonic characteristics are used to measure the degree of harmonic distortion of the load. Identify the transient characteristics of the load, including power mutation point, instantaneous power change rate and dynamic power factor. The power mutation point is used to identify the start and stop state of the load. The instantaneous power change rate represents the dynamic characteristics of the load. The dynamic power factor reflects the change of the power factor of the load at different time points. Step S3, for load characteristics, the maximum relevance minimum redundancy method is used to screen the characteristic data; In step S3, the filtered feature data is weighted by combining the adaptive feature enhancement method, and the feature data in the low signal-to-noise ratio environment is weighted by using the attention mechanism; For load characteristics, the steps of screening characteristic data using the maximum correlation minimum redundancy method are as follows: For each load characteristic , calculate its correlation with the category label The mutual information of is calculated as: , in, Indicates Load characteristics, Indicates the load category, Representation characteristics With category The mutual information between Features Possible values, For Category Possible values, Features The value is And the category is The joint probability of and Respectively represent characteristics and categories The marginal probability of For all feature pairs , calculate mutual information, the formula is: , in, Indicates Load characteristics, Representation characteristics and The mutual information between Features Possible values, Features The value is And features The value is The joint probability of Calculate features Overall rating: , in, Representation characteristics The overall rating of is the selected feature set, is the number of selected features, according to Sort in descending order and select the first Features are used as the final screening results; The filtered features are weighted, and the weights By feature importance index Normalized, the normalized formula is: , Finally, the weighted features are calculated: , in, is the final filtered and weighted feature set, Specifically, the maximum relevance minimum redundancy (mRMR) is used here to screen the load features; first, the mutual information between each feature and the category is calculated to measure the classification relevance of the feature; then the mutual information between features is calculated to quantify the redundancy between features; based on these two indicators, a feature scoring function is constructed so that the final selected feature set can maximize the category information while minimizing the redundant interference between features; During the screening process, the normalization method is used to calculate the feature weights to further optimize the results of feature selection; the screened feature set not only reduces the data dimension and improves the calculation efficiency, but also enhances the generalization ability of the model; on this basis, an adaptive feature enhancement mechanism is introduced to assign weights to features of different importance, thereby improving the classification accuracy and robustness; Step S4, based on the feature data processed in step S3, inputting into a classification model, using a convolutional neural network (CNN) combined with a long short-term memory (LSTM) network to perform load classification; In step S4, CNN is used to extract spatial features and LSTM is used to extract temporal features; In step S4, the model parameters are adjusted in combination with adversarial training to reduce the impact of signal interference on the classification results; Based on the feature data processed in step S3, the classification model is input, and the steps of using convolutional neural network CNN combined with long short-term memory network LSTM to classify the load are as follows: Constructing feature matrix As input to the CNN-LSTM model: , in, Represents the input feature matrix, containing The feature sequence of time steps, Represents the time step The load characteristic vector at Use a one-dimensional convolution kernel to perform convolution operation on the input features: , in, is the feature map after convolution, The convolution kernels are located at The output at For the The convolution kernel weights, is the input feature matrix at position The value at For the The bias term of the convolution kernel, is the size of the convolution kernel, is the activation function, Perform pooling operation: , in, is the feature vector after pooling, and the maximum pooling operation is used to extract information. LSTM calculates the current time step The hidden state of: , in, is the time step The hidden state of is the activation function of the LSTM unit, and are the weight matrices for input and state, respectively, is the hidden state of the previous time step, is the bias term of the LSTM unit, Use the adversarial sample to perturb the input and calculate the perturbed input. The formula is: , in, For adversarial samples, is the disturbance intensity coefficient, is the loss function, Represents the loss function For input The gradient of represents the symbolic function, Finally, CNN-LSTM generates load classification results ; Specifically, in step S4, the CNN-LSTM structure is used to classify the load features; CNN extracts the spatial pattern of load features through one-dimensional convolution operations, and uses the pooling layer to reduce data redundancy and enhance the ability to express local information; then LSTM further captures the temporal dependencies in the time series, so that the model can effectively learn the temporal dynamic characteristics of load changes.

[0027] In addition, in order to improve the robustness of the model to noise interference, an adversarial training strategy is introduced to add small perturbations to generate adversarial samples, so that the model can learn more anti-interference feature representations during training, thereby reducing the occurrence of misclassification. Make full use of the spatiotemporal characteristics of load data to improve the recognition ability of the classification model; CNN is responsible for capturing local patterns of features, while LSTM processes long-term dependencies, so that the final classification result can take into account both short-term and long-term characteristic changes of the load; Step S5, calculating the classification confidence according to the load category in step S4, performing secondary discrimination on low-confidence samples, and adjusting the classification result in combination with the dynamic matching degree; The steps of calculating the classification confidence according to the load category in step S4, performing secondary discrimination on low confidence samples, and adjusting the classification result in combination with the dynamic matching degree are as follows: Calculate the classification confidence, and the classification model outputs the category probability vector : , in, is the classification probability vector, containing The predicted probability of each category, Indicates The predicted probability of the class, Calculating classification confidence : , in, is the confidence level, For low confidence samples, calculate their similarity with the nearest neighbor samples : , in, The current sample and The similarity of neighbor samples is The nearest neighbor sample The eigenvector at If the similarity If it is less than the set threshold, the classification label is adjusted; Calculate the category match: , in, For Category The matching degree, if If it is lower than the set threshold, the classification result will be corrected; Specifically, in step S5, the reliability of the load category output by CNN-LSTM is evaluated by calculating the classification confidence; The classification confidence is calculated based on the probability vector output by the model, which reflects the model's confidence in the current prediction result. For low-confidence samples, the nearest neighbor similarity metric is further introduced to calculate the distance between it and the classified samples to determine whether the current classification is reasonable.

[0028] In addition, the classification results are adjusted in combination with the dynamic matching degree. The dynamic matching degree measures the relative matching degree between the current sample and each category. When the matching degree is lower than the set threshold, the classification result is corrected to improve the reliability of the classification. In practical applications, due to the complexity of load signals, some samples may be difficult to be accurately classified. Therefore, secondary discrimination and dynamic matching strategies are adopted to make the final classification results more consistent with the load operation characteristics and improve the overall load identification ability; Step S6, based on the adjusted load category classification result, establish a load status record, perform operation mode analysis, and monitor abnormal load status; Based on the adjusted load category classification results, establish load status records and conduct operation mode analysis. The steps for monitoring abnormal load status are as follows: According to the load classification result adjusted in step S5, the operating status of the load is recorded and a time series data storage is established. Perform pattern analysis on the recorded data to identify the typical operating modes of the load, including steady-state operation, intermittent operation and sudden changes. Statistical methods and machine learning models are used to detect abnormal load conditions, including power mutations and operating time abnormalities, and historical data is combined to classify abnormalities. Analyze the detected abnormal loads in combination with expert rules and trigger abnormal alarms or adjust classification strategies; Specifically, in step S6, based on the load category after classification adjustment, a load status record is established and an operation mode analysis is performed; the load status record includes information such as load category and load change trend in different time periods. In addition, time series analysis and machine learning models are used to identify typical operating modes of loads, including stable operation, intermittent operation, and abnormal fluctuation modes. For abnormal load conditions, an alarm mechanism is introduced to identify possible abnormal events in real time through threshold detection or deep learning anomaly detection algorithms, and provide corresponding early warning information. The intelligence level of the load monitoring system is improved so that it can not only accurately classify loads, but also further analyze the operating characteristics and abnormal conditions of the loads.

[0029] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A non-intrusive power load identification method, characterized in that: include, Step S1, collecting voltage signals and current signals, obtaining corresponding waveform data, and performing normalization and removing DC offset; Step S2, extracting load characteristics based on the voltage signal and current signal preprocessed in step S1; Step S3, for load characteristics, the maximum relevance minimum redundancy method is used to screen the characteristic data; Step S4, based on the feature data processed in step S3, inputting into a classification model, using a convolutional neural network (CNN) combined with a long short-term memory (LSTM) network to perform load classification; In step S4, CNN is used to extract spatial features and LSTM is used to extract temporal features; In step S4, the model parameters are adjusted in combination with adversarial training; Step S5, calculating the classification confidence according to the load category in step S4, performing secondary discrimination on low-confidence samples, and adjusting the classification result in combination with the dynamic matching degree; Step S6, based on the adjusted load category classification result, establish a load status record, perform operation mode analysis, and monitor abnormal load status.

2. A non-intrusive power load identification method as claimed in claim 1, characterized in that: In step S1, adaptive filtering is used to suppress noise on the current signal to remove background noise; In step S1, the voltage signal and the current signal are decomposed by using a modal decomposition method to extract the target load signal and suppress the interference signal.

3. A non-intrusive power load identification method as claimed in claim 2, characterized in that: In step S1, normalization and DC offset removal are performed, specifically: Collect voltage and current signals and perform preprocessing. Calculate the average value of the acquired signal and subtract it from the original signal. Calculate the maximum amplitude of the signal and divide all data points of the signal by the maximum amplitude to normalize the signal amplitude range to the standardized interval.

4. A non-intrusive power load identification method as claimed in claim 3, characterized in that: The load characteristics include steady-state characteristics and transient characteristics. Steady-state characteristics include power factor, active power, reactive power and harmonic characteristics. Transient characteristics include power mutation point, instantaneous power change rate and dynamic power factor.

5. A non-intrusive power load identification method as claimed in claim 4, characterized in that: The step of extracting load characteristics is: Calculate the steady-state characteristics of the load, including power factor, active power, reactive power and harmonic characteristics. The power factor reflects the phase relationship between current and voltage, the active power represents the energy actually consumed by the load, the reactive power measures the ineffective energy exchange in the power grid, and the harmonic characteristics are used to measure the degree of harmonic distortion of the load. Identify the transient characteristics of the load, including power mutation point, instantaneous power change rate and dynamic power factor. The power mutation point is used to identify the start and stop status of the load. The instantaneous power change rate represents the dynamic characteristics of the load. The dynamic power factor reflects the change of the power factor of the load at different time points.

6. A non-intrusive power load identification method as claimed in claim 5, characterized in that: In step S3, the filtered feature data is weighted calculated in combination with the adaptive feature enhancement method, and the feature data in a low signal-to-noise ratio environment is weighted processed using the attention mechanism.

7. A non-intrusive power load identification method as claimed in claim 6, characterized in that: The step of screening the characteristic data using the maximum relevance minimum redundancy method for the load characteristics is: For each load characteristic , calculate its correlation with the category label The mutual information of is calculated as: ， in, Indicates Load characteristics, Indicates the load category, Representation characteristics With category The mutual information between Features Possible values, For Category Possible values, Features The value is And the category is The joint probability of and Respectively represent characteristics and categories The marginal probability of For all feature pairs , calculate mutual information, the formula is: ， in, Indicates Load characteristics, Representation characteristics and The mutual information between Features Possible values, Features The value is And features The value is The joint probability of Calculate features Overall rating: ， in, Representation characteristics The overall rating of is the selected feature set, is the number of selected features, according to Sort in descending order and select the first Features are used as the final screening results; The filtered features are weighted, and the weights By feature importance index Normalized, the normalized formula is: ， Finally, the weighted features are calculated: ， in, It is the final filtered and weighted feature set.

8. A non-intrusive power load identification method as claimed in claim 7, characterized in that: The step of inputting the classification model based on the feature data processed in step S3 and using the convolutional neural network CNN combined with the long short-term memory network LSTM to classify the load is as follows: Constructing feature matrix As input to the CNN-LSTM model: ， in, Represents the input feature matrix, containing The feature sequence of time steps, Represents the time step The load characteristic vector at Use a one-dimensional convolution kernel to perform convolution operation on the input features: ， in, is the feature map after convolution, The convolution kernels are located at The output at For the The convolution kernel weights, is the input feature matrix at position The value at For the The bias term of the convolution kernel, is the size of the convolution kernel, is the activation function, Perform pooling operation: ， in, is the feature vector after pooling, and the maximum pooling operation is used to extract information. LSTM calculates the current time step The hidden state of: ， in, is the time step The hidden state of is the activation function of the LSTM unit, and are the weight matrices for input and state, respectively, is the hidden state of the previous time step, is the bias term of the LSTM unit, Use the adversarial sample to perturb the input and calculate the perturbed input. The formula is: ， in, For adversarial samples, is the disturbance intensity coefficient, is the loss function, Represents the loss function For input The gradient of represents the symbolic function, Finally, CNN-LSTM generates load classification results .

9. A non-intrusive power load identification method as claimed in claim 8, characterized in that: The steps of calculating the classification confidence according to the load category in step S4, performing secondary discrimination on low confidence samples, and adjusting the classification result in combination with the dynamic matching degree are as follows: Calculate the classification confidence, and the classification model outputs the category probability vector : ， in, is the classification probability vector, containing The predicted probability of each category, Indicates The predicted probability of the class, Calculating classification confidence : ， in, is the confidence level, For low confidence samples, calculate their similarity with the nearest neighbor samples : ， in, The current sample and The similarity of neighbor samples is The nearest neighbor sample The eigenvector at If the similarity If it is less than the set threshold, the classification label is adjusted; Calculate the category match: ， in, For Category The matching degree, if If it is lower than the set threshold, the classification result will be corrected.

10. A non-intrusive power load identification method as claimed in claim 9, characterized in that: The steps of establishing a load status record based on the adjusted load category classification result, performing operation mode analysis, and monitoring abnormal load status are as follows: According to the load classification result adjusted in step S5, the operating status of the load is recorded and a time series data storage is established. Perform pattern analysis on the recorded data to identify the typical operating modes of the load, including steady-state operation, intermittent operation and sudden changes. Statistical methods and machine learning models are used to detect abnormal load conditions, including power mutations and operating time abnormalities, and historical data is combined to classify abnormalities. The detected abnormal load is analyzed in combination with expert rules, and abnormal alarms are triggered or classification strategies are adjusted.

Images