Non-intrusive load behavior refined identification method based on heterogeneous feature fusion

Abstract

The invention discloses a heterogeneous feature fusion-based non-intrusive load behavior fine identification method, which comprises the following steps of: extracting V-I track features based on detection; extracting statistical characteristics; extracting time sequence features; designing a feature fusion model based on the V-I trajectory features, the statistical features and the time sequence features; and performing electric appliance identification based on the feature fusion model. The method improves the recognition accuracy, can extract the multi-dimensional feature information of the electric appliance through the fusion of the V-I track features, the statistical features and the time sequence features, reduces the overlapping of the electric appliance features, and improves the recognition accuracy of the electric appliance. According to the method, the robustness is enhanced, the introduction of the statistical characteristics and the time sequence characteristics can make up for the deficiency of the V-I track characteristics in energy consumption information, the resistance of the system to noise and interference is enhanced, and the stability of electric appliance identification is improved. Compared with a traditional deep learning feature fusion method, the comprehensive performance of the load recognition model can be remarkably improved, and the method is more suitable for actual scenes.

Description

Technical Field

[0001] The present invention belongs to the field of power load monitoring, and specifically relates to a non-invasive load behavior refined identification method based on heterogeneous feature fusion. Background Art

[0002] As global energy demand continues to grow, improving the energy efficiency of power systems has become a key research topic in the energy sector. Non-Intrusive Load Monitoring (NILM) technology, a key component of smart grids, can decompose energy consumption information for individual loads from aggregate power signals, providing users with refined energy consumption data. This helps optimize energy management, reduce power waste, and improve energy efficiency in smart homes and industrial sectors.

[0003] NILM technology was first proposed by Hart in 1992. Its core concept is to leverage the characteristic information in the total power consumption signal, combined with signal processing and pattern recognition methods, to infer the operating status and energy consumption of each independent load. Traditional NILM methods primarily rely on active power (P) and reactive power (Q) features, using cluster analysis and other methods for load identification. However, with the increasing variety of household appliances, these PQ space-based NILM methods have shown limitations when faced with feature overlap, especially for multi-state loads (such as air conditioners, refrigerators, and washing machines) and similar loads (such as fans and ventilation systems).

[0004] In recent years, researchers have begun to explore richer load characteristics, such as current waveforms, harmonic components, and voltage-current (VI) trajectories, to improve the accuracy of load identification. Among them, VI trajectory features have received widespread attention because they can provide information about the nonlinear characteristics and transient behavior of the load, which helps improve the NILM system's ability to identify complex loads. However, relying solely on VI trajectories still has the problem of missing information, such as being unable to reflect the energy consumption level of the load. Therefore, how to integrate VI trajectories with other statistical features to construct a more comprehensive load feature description is an important challenge currently facing the NILM field.

[0005] Current NILM research mainly focuses on two categories of methods: event-based NILM and event-free NILM.

[0006] Event-based methods rely on the detection of load switching events, such as using sudden changes in current or power to signal the on / off state of a load. At high sampling rates, these methods can accurately capture dynamic load behavior, but at low sampling rates, they can easily miss short-duration switching events. The accuracy of event detection directly impacts NILM recognition performance, while noise interference and simultaneous changes in multiple loads can reduce detection reliability.

[0007] The VI trajectory is a graphical load feature that intuitively reflects the load's operating characteristics. Research has shown that different load types have different VI trajectory shapes, making them classifiable using computer vision techniques (such as CNN). However, the construction of VI trajectories relies on sampled voltage and current data and has certain limitations in the following situations:

[0008] Missing information: The VI trace only reflects the relationship between voltage and current and lacks energy information such as power and harmonics.

[0009] Feature overlap: For some loads, such as the ventilation patterns of fans and air conditioners, their VI trajectory shapes may be highly similar, leading to identification confusion.

[0010] To make up for the shortcomings of VI trajectories, researchers began to explore statistical features that complement VI trajectories, such as:

[0011] G1: Amplitude information (active power, reactive power, apparent power, power factor)

[0012] G2: Waveform envelope information (RMS value, skewness, kurtosis, quantile, mean, etc.)

[0013] G3: Frequency domain information (harmonic characteristics, total harmonic distortion THD)

[0014] These statistical features can provide information about the load's overall power consumption and harmonic distribution, complementing the VI trace. For example, when there is significant feature overlap, frequency domain information can improve load differentiation. Therefore, effectively integrating VI traces with statistical features is a key area of ​​current NILM research. Summary of the Invention

[0015] The present invention is proposed to solve the problems existing in the prior art, and its purpose is to provide a non-intrusive load behavior refined identification method based on heterogeneous feature fusion.

[0016] The technical solution of the present invention is: a non-intrusive load behavior refined identification method based on heterogeneous feature fusion, comprising the following steps:

[0017] A. Extract VI trajectory features based on detection;

[0018] B. Extract statistical features;

[0019] C. Extracting temporal features;

[0020] D. Design a feature fusion model based on VI trajectory features, statistical features, and time series features;

[0021] E. Electrical appliance identification based on feature fusion model.

[0022] Furthermore, step A extracts VI trajectory features based on detection. The specific process is as follows:

[0023] First, detect the switching event of the appliance;

[0024] Then, for the switching events, the voltage and current data of the electrical appliances are extracted;

[0025] Finally, VI trajectory features are constructed based on the extracted voltage and current data.

[0026] Furthermore, the VI trace feature converts voltage and current data into image representation, providing multi-dimensional electrical appliance feature information.

[0027] Furthermore, step B extracts statistical features. The specific process is as follows:

[0028] First, extract the time domain statistical characteristics of the electrical appliance;

[0029] Then, the frequency domain statistical characteristics of the electrical appliance are extracted;

[0030] Finally, the physical information of the electrical appliance is provided through time domain statistical features and frequency domain statistical features, which makes up for the deficiency of VI trajectory features in energy consumption information.

[0031] Furthermore, the time domain statistical features and frequency domain statistical features include active power, reactive power, power factor, and harmonic current features.

[0032] Furthermore, step C extracts time series features. The specific process is as follows:

[0033] First, the load current signal is normalized to improve the versatility and numerical stability of the model;

[0034] Then, the wavelet packet decomposition method is used to perform multi-scale decomposition on the current signal to extract multi-dimensional sub-band signals reflecting the energy distribution of different frequency bands;

[0035] Finally, the decomposed multidimensional sequence is input into the long short-term memory (LSTM) network to extract the dynamic change pattern of the electrical appliance in the time domain dimension, which is used to characterize the time dependence and operating rules of the electrical appliance during start-up and shutdown, thereby making up for the shortcomings of image features in terms of temporal expression ability.

[0036] Furthermore, step D designs a feature fusion model based on VI trajectory features, statistical features, and time series features. The specific process is as follows:

[0037] Perform feature splicing and fusion based on VI trajectory features, statistical features, and time series features;

[0038] Furthermore, the feature splicing and fusion simultaneously integrates VI trajectory features, statistical features, and time series features during the deep learning process. The specific process is as follows:

[0039] d11. Use the VI model, statistical model, and time series model to extract the latent feature vectors of VI trajectory features, statistical features, and time series features respectively;

[0040] d12. Concatenate the three obtained feature vectors to form a fused feature vector;

[0041] d13. Perform nonlinear transformation on the fused feature vector through the fully connected layer to generate the final appliance recognition result.

[0042] The beneficial effects of the present invention are as follows:

[0043] The present invention improves recognition accuracy and can extract multi-dimensional feature information of electrical appliances by fusing VI trajectory features, statistical features and time series features, thereby reducing the overlap of electrical appliance features and improving the accuracy of electrical appliance recognition.

[0044] The present invention enhances robustness, and the introduction of statistical features and time series features can make up for the deficiency of VI trajectory features in energy consumption information, enhance the system's resistance to noise and interference, and improve the stability of electrical appliance identification.

[0045] The present invention reduces computational costs. Compared with traditional deep learning methods, the feature fusion method proposed in the present invention significantly reduces computational costs while maintaining high performance, and is suitable for real-time application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is the load voltage and current waveform extraction module based on events in the present invention;

[0047] Figure 2 It is the module for extracting, fusing and identifying load heterogeneous features in the present invention. DETAILED DESCRIPTION

[0048] Hereinafter, the present invention will be described in detail with reference to the accompanying drawings and embodiments:

[0049] like Figure 1 As shown in FIG, the non-intrusive load behavior fine-grained identification method based on heterogeneous feature fusion includes the following steps:

[0050] A. Extract VI trajectory features based on detection;

[0051] B. Extract statistical features;

[0052] C. Extracting temporal features;

[0053] D. Design a feature fusion model based on VI trajectory features, statistical features, and time series features;

[0054] E. Electrical appliance identification based on feature fusion model.

[0055] Step A extracts VI trajectory features based on detection. The specific process is as follows:

[0056] First, detect the switching event of the appliance;

[0057] Then, for the switching events, the voltage and current data of the electrical appliances are extracted;

[0058] Finally, VI trajectory features are constructed based on the extracted voltage and current data.

[0059] The VI trace feature converts voltage and current data into image representation, providing multi-dimensional electrical appliance feature information.

[0060] Step B extracts statistical features. The specific process is as follows:

[0061] First, extract the time domain statistical characteristics of the electrical appliance;

[0062] Then, the frequency domain statistical characteristics of the electrical appliance are extracted;

[0063] Finally, the physical information of the electrical appliance is provided through time domain statistical features and frequency domain statistical features, which makes up for the deficiency of VI trajectory features in energy consumption information.

[0064] The time domain statistical features and frequency domain statistical features include active power, reactive power, power factor, and harmonic current features.

[0065] Step C extracts time series features. The specific process is as follows:

[0066] First, the load current signal is normalized to improve the versatility and numerical stability of the model;

[0067] Then, the wavelet packet decomposition method is used to perform multi-scale decomposition on the current signal to extract multi-dimensional sub-band signals reflecting the energy distribution of different frequency bands;

[0068] Finally, the decomposed multidimensional sequence is input into the long short-term memory (LSTM) network to extract the dynamic change pattern of the electrical appliance in the time domain dimension, which is used to characterize the time dependence and operating rules of the electrical appliance during start-up and shutdown, thereby making up for the shortcomings of image features in terms of temporal expression ability.

[0069] Step D designs a feature fusion model based on VI trajectory features, statistical features, and time series features. The specific process is as follows:

[0070] Perform feature splicing and fusion based on VI trajectory features, statistical features, and time series features;

[0071] The feature splicing and fusion process splices and fuses the learned VI trajectory features, statistical features, and time series feature vectors during the deep learning process. The specific process is as follows:

[0072] d11. Use the VI model, statistical model, and time series model to extract the latent feature vectors of VI trajectory features, statistical features, and time series features respectively;

[0073] d12. Concatenate the three obtained feature vectors to form a fused feature vector;

[0074] d13. Perform nonlinear transformation on the fused feature vector through the fully connected layer to generate the final appliance recognition result.

[0075] Specifically, step E performs electrical appliance identification based on the feature fusion model, as follows:

[0076] Based on the fused feature information, a deep learning model is used to classify and identify electrical appliances.

[0077] Experimental verification shows that the feature fusion method proposed in this invention demonstrates excellent electrical appliance recognition performance on multiple public datasets.

[0078] Example 1

[0079] The non-intrusive load behavior fine-grained identification method based on heterogeneous feature fusion includes the following steps:

[0080] A. Extract VI trajectory features based on detection

[0081] a1. Perform data preprocessing

[0082] The collected voltage and current data are preprocessed, including outlier processing and missing value repair. Common methods include median filtering, linear interpolation, and Hermite interpolation.

[0083] a2. Switch event detection

[0084] Assuming that only one electrical appliance switches state at a time, the switching event of the appliance can be detected by analyzing the changes in power or current data.

[0085] a3. Data segmentation

[0086] After detecting the switching event, the voltage and current data before and after the event are extracted and expressed as V event―start 、V event―end , I event―start and I event―end .

[0087] a4.VI trajectory construction

[0088] By calculating the phase angles of the voltage and current waveforms, the initial sampling point with zero phase angle is selected to ensure that the current waveforms can be directly subtracted.

[0089] The voltage and current waveforms are expressed as:

[0090]

[0091] I=I event―end ―I event―start

[0092] a5.VI trajectory visualization

[0093] The normalized voltage and current waveforms are plotted as VI traces and encoded using the HSV color space.

[0094] B. Extracting statistical features

[0095] b1. Time domain feature extraction

[0096] Extract time domain features of electrical appliances such as active power, reactive power, power factor, apparent power, RMS value, percentiles (25% and 75%), skewness, kurtosis, mean, minimum, maximum, median, variance, standard deviation and energy.

[0097] The calculation formulas for skewness and kurtosis are:

[0098]

[0099] Where μ is the mean, σ is the standard deviation, and T is the number of data points.

[0100] b2. Frequency domain feature extraction

[0101] Extract the harmonic current characteristics (1st, 3rd, 5th, 7th, 9th, 11th, 13th, 15th harmonic values) and total harmonic distortion (THD) of electrical appliances.

[0102] The calculation formula for THD is:

[0103]

[0104] Among them, H1 is the amplitude of the fundamental component, H2, H3, H n is the amplitude of the 2nd, 3rd to nth harmonics.

[0105] C. Extracting temporal features

[0106] c1. Original signal selection and preprocessing

[0107] The input signal is the current steady-state waveform I(t) extracted after event detection, with a length of T. I(t) is normalized to improve the model's adaptability to electrical appliances with different amplitudes:

[0108]

[0109] c2. Wavelet Packet Decomposition (WPD)

[0110] Use DB1 wavelet basis to normalize signal I norm (T) Perform three-layer wavelet packet decomposition. Unlike traditional wavelet decomposition that only processes low frequencies, WPD decomposes both low-frequency and high-frequency signals, thereby obtaining finer-grained time-frequency features. The decomposition process is as follows:

[0111]

[0112] Among them, h(k) and g(k) are the low-pass and high-pass filter coefficients respectively, and u is the current decomposition level. The characteristic matrix composed of each sub-signal is R n , where n is the sample length.

[0113] c3. Long Short-Term Memory (LSTM) Modeling

[0114] A two-layer LSTM network is designed to model the temporal correlation and long-term dependency between sub-band signals. The number of hidden units in the first layer of LSTM is H1, and the number in the second layer is H2. Each time, a hidden state vector is output. The final output sequence is mapped to a fixed-length embedding vector f through a fully connected seq

[0115]

[0116] c4. Feature normalization output

[0117] Output time series feature vector f seq , as the input for subsequent feature fusion. To improve numerical stability and model convergence, Batch Normalization is performed on the feature vector.

[0118] D. Design feature fusion model based on VI trajectory features, statistical features, and time series features

[0119] d1. To evaluate the potential improvement of fusing VI and statistical features for instrument identification, baseline methods were established using VI trajectory features and statistical features, respectively.

[0120] d2. Different methods uniformly use the same classification network architecture as a comparative benchmark for evaluating the performance of various feature fusion methods to ensure fairness and comparability of the results.

[0121] Among them, the VI image feature channel uses the improved ResNet18 as the classification backbone network. Its structure is lightweight while maintaining the advantages of the original residual connection to adapt to practical application requirements such as the limited number of image data samples and limited model deployment resources in non-invasive load monitoring.

[0122] The input VI trajectory image is first resized to a 224×224×3 color image through interpolation resampling. This size is compatible with mainstream convolutional neural network architectures and helps fully preserve the spatial distribution and texture characteristics of the voltage-current trajectory. Subsequently, the image features are gradually abstracted and compressed through multiple convolutional modules, residual blocks, and fully connected layers, ultimately outputting the hidden vector of the device.

[0123] d3. Considering that statistical features are typically global descriptors extracted from load signals, such as mean, variance, energy, kurtosis, and harmonic components, these features are dimensionally independent and lack significant temporal and spatial correlation. Therefore, they are not suitable for processing using complex networks with time series modeling capabilities or spatial convolutional structures. To this end, this paper uses a simple feedforward neural network consisting of two fully connected layers to model the statistical features, as follows:

[0124] d31. All extracted time-domain and frequency-domain statistical features are concatenated into a one-dimensional vector as the network input. After being mapped through two fully connected layers containing nonlinear activation functions (such as ReLU or LeakyReLU), a feature representation of uniform dimension is formed.

[0125] d32. Through an output layer, high-dimensional features are compressed into predicted probability distributions for appliance categories. This lightweight and efficient structure fully exploits the discriminative information in statistical features while maintaining low computational overhead. This network architecture, designed specifically for the structural characteristics of statistical features, effectively avoids overfitting and improves the model's recognition stability and generalization capabilities in scenarios with small sample sizes or limited resources.

[0126] d4. Considering that the original load current signal has significant sequential dependence and dynamic change characteristics in the time dimension, the present invention designs a time series feature extraction structure based on the long short-term memory network (LSTM), such as Figure 2 As shown, the details are as follows:

[0127] d41. The original current signal is processed by three-layer wavelet packet decomposition to obtain a sequence feature matrix containing multi-scale frequency domain information.

[0128] This feature matrix is ​​fed into a sequence modeling structure consisting of two stacked LSTM networks to capture long-term dependencies and temporal patterns between sub-signals. Based on the final hidden state of the LSTM output, a fully connected layer is added to perform feature compression and nonlinear mapping, resulting in a feature vector of uniform length.

[0129] d43. After being processed by the activation function, this feature vector serves as the feature output of the time series channel and participates in the subsequent bimodal feature fusion. This structure fully leverages the advantages of LSTM in processing time series data and can effectively model the dynamic evolution characteristics of current signals in different frequency subbands, thereby improving the load identification model's ability to express and discriminate complex electricity consumption behaviors.

[0130] For the feature fusion model, a feature concatenation fusion method is used to simultaneously fuse VI trajectory features, statistical features, and timing feature information. It first learns VI trajectory features, statistical features, and timing features independently, and then concatenates the feature vectors learned by each module. This combined vector is then used for further learning to generate device prediction results, effectively utilizing the combined information to improve accuracy, such as Figure 2 shown.

[0131] The above approach enables the model to capture and understand the complex interactions among VI trace features, statistical features, and temporal features.

[0132] E. Appliance identification based on feature fusion model

[0133] The system architecture for this non-intrusive load behavior refinement identification method, which utilizes heterogeneous feature fusion, is based on non-intrusive load monitoring (NILM) technology. This approach aims to improve the accuracy and robustness of appliance identification by integrating VI trajectory features, statistical features, and temporal features. The system architecture primarily includes the following modules: data acquisition and preprocessing, feature extraction, feature fusion, and appliance identification.

[0134] Specifically, the data acquisition and preprocessing module is responsible for collecting voltage and current data and preprocessing the data, including outlier processing, missing value repair, and data segmentation.

[0135] Specifically, the feature extraction module extracts VI trajectory features, statistical features, and time series features from the preprocessed data.

[0136] Specifically, the feature fusion module fuses VI trajectory features, statistical features and temporal features according to different fusion strategies.

[0137] Specifically, the appliance recognition module uses a deep learning model to classify and identify appliances based on the fused feature information.

[0138] Specifically, data acquisition in the data acquisition and preprocessing module is the first step in the NILM system. Voltage and current data are typically collected through smart meters or high-precision sensors. The collected data may contain outliers or missing values, so preprocessing is required.

[0139] The main steps of preprocessing are as follows:

[0140] Outlier processing: Use methods such as median filtering to remove outliers in the data to ensure data accuracy.

[0141] Missing value repair: For missing voltage or current data, linear interpolation or Hermite interpolation is used to repair it.

[0142] Data segmentation, by detecting the switching events of electrical appliances, the voltage and current data are segmented, and the data before and after the event are extracted for subsequent feature extraction.

[0143] Specifically, the VI trajectory feature in the feature extraction module is one of the core features of the present invention, which can intuitively reflect the operating status of the appliance. The extraction of the VI trajectory feature is as follows:

[0144] Switching event detection assumes that only one electrical appliance switches state at a time. By analyzing changes in power or current data, the switching event of the appliance is detected.

[0145] Data segmentation,After detecting a switching event, the voltage and current data before and after the event are,extracted and represented as the voltage and current waveforms at the beginning and,end of the event, respectively.

[0146] The VI trace is constructed by calculating the phase angles of the voltage and current waveforms and selecting the initial sampling point with zero phase angle to ensure that the current waveforms can be directly subtracted. The normalized voltage and current waveforms are plotted as VI traces and encoded using the HSV color space.

[0147] Specifically, the statistical features in the feature extraction module are another core feature of the present invention, which can provide physical information of the electrical appliance and make up for the deficiency of the VI trajectory feature in energy consumption information.

[0148] The extraction of statistical features includes time domain features and frequency domain features, as follows:

[0149] Time domain features, extracting time domain features such as active power, reactive power, power factor, apparent power, root mean square value, percentiles (25% and 75%), skewness, kurtosis, mean, minimum, maximum, median, variance, standard deviation and energy of electrical appliances.

[0150] Frequency domain features extract the harmonic current characteristics (1st, 3rd, 5th, 7th, 9th, 11th, 13th, and 15th harmonic values) and total harmonic distortion (THD) of electrical appliances. Frequency domain features can capture the periodicity and harmonic characteristics of electrical appliance signals, helping to distinguish different types of appliances.

[0151] The extraction of timing features is based on the dynamic change characteristics of the electrical current signal, as follows:

[0152] Firstly, the original current signal is decomposed into multiple scales using wavelet packet decomposition (WPD) to obtain sub-band signals of different frequency bands, thereby preserving the changing details of the electrical appliance in the time-frequency domain.

[0153] The multidimensional sequence composed of these subband signals is then fed into a two-layer long short-term memory (LSTM) network. Leveraging its ability to model temporal dependencies, the LSTM extracts time series feature vectors that reflect the appliance's start-stop behavior and operating mode. These features capture the signal's temporal trends, fluctuation cycles, and transient characteristics, helping to characterize the dynamic response of appliance loads over time and thus improving the ability to distinguish between different appliances.

[0154] Specifically, the feature fusion module uses feature splicing and fusion.

[0155] Specifically, the appliance recognition module uses a deep learning model to classify and identify appliances based on the fused feature information, as follows:

[0156] The VI model uses an improved ResNet-18 architecture to process VI trajectory images and extract feature vectors. A fully connected layer maps the feature vectors to the prediction space to generate predicted probabilities for the appliance.

[0157] The statistical model uses a simple fully connected neural network to process statistical features and extract feature vectors. The output layer generates the predicted probability of the appliance.

[0158] Time series model,The time series model uses a two-layer long short-term memory network (LSTM) to model the decomposed current sequence, extract the time series feature vector reflecting the dynamic behavior of the electrical appliance, and generate the prediction probability of the corresponding electrical appliance through the output layer.

[0159] The fusion model, according to the feature splicing and fusion strategy, fuses the VI trajectory features, statistical features and timing features, and generates the final appliance recognition results through the fully connected layer.

[0160] Specifically, the appliance recognition module is implemented based on the PyTorch deep learning framework and uses a five-fold cross-validation method for model training and evaluation to prevent model overfitting. The details are as follows:

[0161] Model training: Initialize the model with random weights and train with the same optimizer and learning rate. During training, cross-validation is used to ensure the generalization ability of the model.

[0162] Hyperparameter tuning: perform sensitivity analysis on the model's hyperparameters (such as learning rate and batch size) and select the optimal hyperparameter combination to improve model performance.

[0163] Performance evaluation, using indicators such as precision, recall, and F1 score (F-score) to evaluate the model's appliance recognition performance.

[0164] The present invention can effectively improve the appliance identification performance of non-intrusive load monitoring systems, and provide strong technical support for energy management in fields such as smart homes, commercial buildings, and industrial monitoring.

[0165] Verification Example

[0166] Dataset preparation

[0167] This paper experimentally validates the proposed method on two public datasets: PLAID and WHITED. The PLAID dataset records voltage and current data for individual appliances at a sampling frequency of 30 kHz, containing 1,793 appliance events spanning 11 appliance types from 82 different brands. The WHITED dataset collects voltage and current measurement data for individual appliances at a sampling frequency of 44.1 kHz, containing 1,339 appliance events spanning 54 appliance types.

[0168] Experimental results

[0169] On the PLAID dataset, the F-macro of the model using only statistical features was 0.8904, while the F-macro of the model using VI trajectory features increased to 0.9613, and the F-macro of the model using time series features was 0.9487. Using a feature fusion strategy, the F-macro of the model reached 0.9780, significantly improving appliance recognition performance.

[0170] On the WHITED dataset, the F-macro of the model using only statistical features was 0.8133, while the F-macro of the model using VI trajectory features increased to 0.9660, and the F-macro of the model using statistical features was 0.9568. Using the feature fusion strategy, the F-macro of the model reached 0.9900, further verifying the effectiveness of the proposed method.

[0171] Experimental results demonstrate that the proposed feature fusion method can effectively improve the accuracy and robustness of appliance identification. By fusing VI trajectory features, statistical features, and temporal features, it extracts multidimensional feature information about appliances, reduces overlap in appliance features, and thus improves appliance identification accuracy. Furthermore, while maintaining high performance, the proposed method significantly reduces computational costs, making it suitable for real-time applications.

Claims

1. A non-intrusive load behavior refinement identification method based on heterogeneous feature fusion, characterized by: The following steps are involved: A. Extract VI trajectory features based on detection; B. Extract statistical features; C. Extracting temporal features D. Design a feature fusion model based on VI trajectory features, statistical features, and time series features; E. Electrical appliance identification based on feature fusion model.

2. The non-intrusive load behavior refined identification method based on heterogeneous feature fusion according to claim 1 is characterized by: Step A extracts VI trajectory features based on detection. The specific process is as follows: First, detect the switching event of the appliance; Then, for the switching events, the voltage and current data of the electrical appliances are extracted; Finally, VI trajectory features are constructed based on the extracted voltage and current data.

3. The non-intrusive load behavior fine-grained identification method based on heterogeneous feature fusion according to claim 2 is characterized by: The VI trace feature converts voltage and current data into image representation, providing multi-dimensional electrical appliance feature information.

4. The non-intrusive load behavior refined identification method based on heterogeneous feature fusion according to claim 1 is characterized by: Step B extracts statistical features. The specific process is as follows: First, extract the time domain statistical characteristics of the electrical appliance; Then, the frequency domain statistical characteristics of the electrical appliance are extracted; Finally, the physical information of the electrical appliance is provided through time domain statistical features and frequency domain statistical features, which makes up for the deficiency of VI trajectory features in energy consumption information.

5. The non-intrusive load behavior refined identification method based on heterogeneous feature fusion according to claim 1 is characterized by: Step C extracts time series features. The specific process is as follows: First, the load current signal is normalized to improve the versatility and numerical stability of the model; Then, the wavelet packet decomposition method is used to perform multi-scale decomposition on the current signal to extract multi-dimensional sub-band signals reflecting the energy distribution of different frequency bands; Finally, the decomposed multidimensional sequence is input into the long short-term memory (LSTM) network to extract the dynamic change pattern of the electrical appliance in the time domain dimension, which is used to characterize the time dependence and operating rules of the electrical appliance during start-up and shutdown, thereby making up for the shortcomings of image features in terms of temporal expression ability.

6. The non-intrusive load behavior refined identification method based on heterogeneous feature fusion according to claim 4 is characterized by: The time domain statistical features and frequency domain statistical features include active power, reactive power, power factor, and harmonic current features.

7. The non-intrusive load behavior refined identification method based on heterogeneous feature fusion according to claim 1 is characterized by: Step D designs a feature fusion model based on VI trajectory features, statistical features, and time series features. The specific process is as follows: Feature splicing and fusion are performed based on VI trajectory features, statistical features, and time series features.

8. The non-intrusive load behavior refined identification method based on heterogeneous feature fusion according to claim 7 is characterized by: The feature splicing and fusion simultaneously integrates VI trajectory features, statistical features, and time series features during the deep learning process. The specific process is as follows: d11. Use the VI model, statistical model, and time series model to extract the latent feature vectors of VI trajectory features, statistical features, and time series features respectively; d12. Concatenate the three obtained feature vectors to form a fused feature vector; d13. Perform nonlinear transformation on the fused feature vector through the fully connected layer to generate the final appliance recognition result.

Images