Non-intrusive load identification method based on load high-frequency feature imaging

Abstract

The invention discloses a non-intrusive load identification method based on load high-frequency feature imaging, and aims to realize comprehensive optimization of economical efficiency, low-carbon property and stability of a power system. At a power supply inlet end, a high-precision constantan resistor and a high-speed data acquisition card are used for acquiring current and voltage data, frequency domain information is acquired through Fourier transform, and related parameters are calculated to complete feature extraction. Converting the frequency domain features into a two-dimensional image through discrete cosine transform, and synthesizing current and power feature images into an RGB image by using a Laplacian pyramid fusion technology; and finally, selecting a lightweight deep learning model OfficientNet, training by using a fused image, and adjusting weight deviation to identify load characteristics. The method overcomes the defects of the traditional technology in the aspect of load identification, facilitates reasonable distribution of electric power resources, reduction of energy consumption, improvement of energy utilization efficiency and guarantee of stable and reliable operation of an electric power system through accurate identification of the load, and has important application value in the field of electric power system management.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent monitoring and analysis of power systems, and specifically relates to a non-intrusive load identification method based on imaging of high-frequency characteristics of loads. Background Art

[0002] With the continuous growth of electricity demand and the increasing complexity of power systems, it is crucial to accurately understand the composition and characteristics of power loads. Traditional intrusive load monitoring methods require the installation of sensors on each load device, which is not only costly, but also cumbersome to install and maintain, making it difficult to promote and apply on a large scale. Non-intrusive load identification technology has emerged, but existing non-intrusive methods still have many shortcomings.

[0003] In terms of feature extraction, most traditional technologies rely only on simple time domain or low-frequency domain analysis, such as focusing only on the effective value of current and voltage or simple power calculation, which cannot fully capture the complex characteristics of the load. For some load devices with similar power consumption but completely different working principles and operating modes, such as different types of motors or electronic equipment, these traditional methods are difficult to effectively distinguish.

[0004] In terms of load identification models, the early model structures were relatively simple, such as rule-based expert systems or simple machine learning models, which had limited learning capabilities for complex load characteristics and were difficult to cope with the diverse load types and dynamically changing operating conditions in modern power systems. In addition, with the increase in nonlinear loads in power systems, such as variable frequency speed regulation equipment and electronic rectifiers, the complex signal components such as harmonics and interharmonics generated by them pose greater challenges to traditional identification methods. A more advanced and efficient load identification technology is urgently needed to meet the needs of refined management of power systems. Summary of the invention

[0005] In view of the above deficiencies in the prior art, the purpose of the present invention is to achieve deep optimization of the power system through innovative technical means. By using the high-frequency feature visualization of loads and deep learning models, various types of loads can be accurately identified, the rationality of power distribution can be improved, energy consumption and costs can be reduced, and economy can be achieved; energy conservation and emission reduction can be promoted, low-carbon transformation can be promoted, and low-carbon properties can be achieved; the ability to respond to load changes can be enhanced, the stable operation of the system can be guaranteed, and stability can be ensured, providing a comprehensive solution for the efficient operation of the power system.

[0006] In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0007] The non-intrusive load identification method based on the high-frequency characteristic imaging of the load includes the following steps:

[0008] S1. Data collection and feature extraction. At the power supply inlet, the current and voltage data in the load circuit are collected by connecting a constantan resistor in series. The collected time domain signals are converted into frequency domain signals using Fourier transform. The fundamental frequency and the amplitude and phase of each harmonic are extracted from the frequency domain signals, and the power factor and total harmonic distortion are calculated.

[0009] S2. Feature image processing, converting the frequency domain features into a two-dimensional image through discrete cosine transform (DCT), the two-dimensional image includes the feature information of the current and the feature information of the power, each feature value in the two-dimensional image corresponds to a pixel value in the image, thereby converting the one-dimensional feature vector into a two-dimensional image;

[0010] S3. Feature image fusion, using Laplace pyramid fusion technology, the current feature image and the power feature image are merged into an RGB image. In this process, the current feature image is used as the red channel and the power feature image is used as the green channel. If necessary, other channels can be added;

[0011] S4. Model training: Select a lightweight deep learning model (EfficientNet) and use the fused feature images for training. By adjusting the weights and biases of the model, the model can learn to recognize different load feature images.

[0012] Furthermore, the S1 includes:

[0013] S11. Data acquisition: At the power supply inlet, high-precision, low-temperature-coefficient constantan resistors are used, and their resistance values ​​are strictly calibrated to ensure that the measurement error under different currents is extremely small; a professional-grade high-speed data acquisition card is used, and its sampling frequency is set between 10kHz and 20kHz, which can be flexibly adjusted according to the actual load characteristics;

[0014] S12. Frequency domain conversion and harmonic analysis: Fast Fourier transform is performed on the collected time domain signal. For discrete time signals, the discrete Fourier transform (DFT) formula is:

[0015]

[0016] Where n is the index of the sequence; k is the frequency domain index, which indicates the position of different frequency components in the frequency domain; x(n) is the discrete time signal; N is the length of the discrete sequence; is a complex exponential function, where j is the imaginary unit (j 2 =-1), the complex exponential function plays a key role in mapping the time domain signal to the frequency domain in DFT. Its periodicity and orthogonality make it possible to decompose the time domain signal into a combination of sine and cosine components of different frequencies through this formula;

[0017] The frequency domain signal is calculated by FFT algorithm, and the fundamental frequency and the amplitude and phase of each harmonic are extracted;

[0018] During the extraction process, a peak detection algorithm and a phase calculation function are used to accurately locate the harmonic frequency components and their corresponding amplitude and phase information;

[0019] S13. Power factor and total harmonic distortion calculation,

[0020] First, calculate the active power. The active power is calculated using the numerical integration method. The formula is as follows:

[0021]

[0022] Where T is the period of the current and voltage signals, which reflects the repeatability of the signals in time; M is the number of time intervals after discretization, and Δt is the time interval; u(t) and i(t) are the voltage and current signals that change with time, respectively, and m is the index of the discrete time point; u(mΔt) and i(mΔt) are the voltage and current values ​​at the discrete time point mΔt;

[0023] Then, calculate the apparent power. When calculating the apparent power, first calculate the effective values ​​of voltage and current respectively through the root mean square algorithm.

[0024] Voltage effective value U:

[0025]

[0026] Current effective value I:

[0027]

[0028] Apparent power:

[0029] S = UI;

[0030] Finally, calculate the total harmonic distortion using the formula:

[0031]

[0032] Among them, A n It represents the amplitude of the nth harmonic, A1 is the amplitude of the fundamental wave; the fundamental wave is the main frequency component of the signal, and the harmonics are the frequency components other than the fundamental wave.

[0033] Furthermore, in S2, the frequency domain features extracted by S1 are input into a two-dimensional discrete cosine transform (DCT) module. For a two-dimensional signal, the DCT transform formula is:

[0034]

[0035]

[0036] x(n,m) is the frequency domain feature data obtained by S1; C(u,v) is the result of two-dimensional discrete cosine transform (DCT); u and v are the coordinates in the two-dimensional image; α(v) and u(v) are coefficients;

[0037] Through the above transformation, the one-dimensional frequency domain feature vector is converted into a two-dimensional image, in which each pixel value corresponds to a DCT coefficient, and the row and column structure of the image intuitively presents the spatial distribution characteristics of the frequency domain features, thereby generating a two-dimensional image containing the characteristic information of the current and the characteristic information of the power.

[0038] Furthermore, in S3, the step of fusion of feature images includes:

[0039] S31. Laplacian pyramid construction, constructing Laplacian pyramids for the current feature image and the power feature image respectively;

[0040] Taking the current characteristic image as an example, we first construct a Gaussian pyramid and set the original image as the 0th layer, that is, Then, after Gaussian blurring the current layer image, alternate row and column sampling is performed to obtain the next layer image, that is, Repeat this process to construct L-layer Gaussian pyramid; then calculate the Laplacian pyramid: Expand is an upsampling operation, which starts from the L-1th layer and calculates upward to obtain the Laplacian pyramid representation of the current feature image:

[0041]

[0042] The same method is used to construct the Laplacian pyramid of the power feature image;

[0043] S32. Image fusion: The Laplacian pyramid of the current feature image is used as the red channel of the RGB image; the Laplacian pyramid of the power feature image is used as the green channel. If other feature images need to be added, their Laplacian pyramids can be constructed according to the same process and assigned to the corresponding channels, thereby realizing multi-feature image fusion and generating the Laplacian pyramid representation of the fused image:

[0044]

[0045] S33. Reconstruct the fused image. Reconstruct the image from the lowest level of the fused Laplacian pyramid. First, Upsample to get the upsampled image Then with Add, that is,

[0046] Then The upsampling and addition operations are repeated, and so on, until the final reconstructed RGB fused image is obtained, which completely retains the fused multi-feature information and provides rich input data for subsequent model training.

[0047] Furthermore, the S4 comprises the following steps:

[0048] S41. Data preparation and enhancement: The fused feature image dataset is divided into training set, validation set and test set in a ratio of 7:2:1. Multi-dimensional enhancement operations are performed on the training set data, including random rotation, horizontal and vertical flipping, random cropping and adding Gaussian noise, which greatly expands the diversity of training data and improves the generalization ability of the model.

[0049] S42. Model selection and initialization. According to the complexity of load characteristics and the size of data sets, for relatively simple load scenarios and small data sets, select the appropriate version in the lightweight deep learning model EfficientNet family, and select EfficientNet-B0; for complex loads and large data sets, you can choose EfficientNet-B4, etc.; use the Kaiming initialization method to initialize the weights of the convolutional layer, calculate the appropriate initialization weight variance according to the number of input and output channels of the convolutional layer, and initialize the bias to 0. At the same time, set the initial learning rate between 0.001-0.01, the weight decay coefficient between 0.0001-0.001, the momentum parameter to 0.9 and other hyperparameters to build the initial environment for model training;

[0050] S43. Training and optimization, defining the cross entropy loss function Among them, C is the number of categories; y is the true label; To predict the output, the Adam optimizer is selected, and its parameter update formula is:

[0051] m t =β1m t-1 +(1-β1)g t ;

[0052]

[0053] Among them, β1 = 0.9; β2 = 0.999; η is the learning rate; ε = 1e-8; g t is the current gradient, θ tare model parameters; during the training process, image data is input into the model in batches of 32-128, and forward propagation is performed to calculate the predicted output. After calculating the loss function, the model parameters are updated through back propagation; after each training epoch, the model performance is evaluated on the validation set, and indicators such as accuracy, recall rate, and F1 value are calculated. The hyperparameters are dynamically adjusted according to the validation set results to ensure the effectiveness and stability of model training.

[0054] In summary, due to the adoption of the above technical solution, the beneficial technical effects of the invention are:

[0055] Improve the economic efficiency of the power system: Through accurate load identification, power system operators can more accurately understand the electricity demand and behavior patterns of different loads. This helps to formulate a more reasonable time-of-use electricity price strategy, guide users to optimize electricity consumption time, reduce peak loads and fill valleys, reduce the peak-to-valley difference of the power system, and reduce the demand for spare power generation capacity, thereby saving power generation costs. At the same time, in terms of maintenance and upgrading of power equipment, based on the load identification results of the present invention, high-energy-consuming or critical load equipment can be optimized and transformed in a targeted manner to improve the energy utilization efficiency of the equipment, further reduce operating costs, and achieve economical operation of the power system.

[0056] Enhance the low-carbon nature of the power system: Accurate load identification provides strong support for energy management. By identifying different types of loads, especially those with high energy consumption and low energy efficiency, effective energy-saving measures can be taken, such as energy efficiency evaluation and optimization control of large motors and other equipment in industrial production, to reduce unnecessary energy waste. In addition, in power systems with renewable energy access, accurate load information helps to better coordinate renewable energy generation and load demand, improve the absorption rate of renewable energy, and reduce dependence on traditional fossil energy generation, thereby reducing carbon emissions and promoting the transformation of the power system to a low-carbon one.

[0057] Ensure the stability of the power system: The present invention can monitor the load changes in the power system in real time and promptly discover abnormal load behavior or potential faults. For some large loads that are suddenly connected or withdrawn, as well as nonlinear loads that may cause voltage fluctuations, harmonic pollution and other problems, through accurate identification and monitoring, corresponding control measures can be taken in advance, such as adjusting reactive power compensation devices, optimizing the operation mode of the power grid, etc., to maintain the voltage and frequency stability of the power system, ensure the safe and reliable operation of the power system, improve the power supply quality and stability of the power system, reduce the probability of power outages, and reduce the losses caused by power outages to social production and life.

[0058] In summary, the present invention achieves comprehensive optimization of the economy, low carbon and stability of the power system through innovative technical means, and provides important technical support for the sustainable development of modern power systems. BRIEF DESCRIPTION OF THE DRAWINGS DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the invention more clear, the invention is further described in detail below in combination with the embodiments. It should be understood that the specific embodiments described here are only used to explain the invention and are not used to limit the invention.

[0060] The non-intrusive load identification method based on the high-frequency characteristic imaging of the load includes the following steps:

[0061] S1. Data collection and feature extraction. At the power supply inlet, the current and voltage data in the load circuit are collected by connecting a constantan resistor in series. The collected time domain signals are converted into frequency domain signals using Fourier transform. The fundamental frequency and the amplitude and phase of each harmonic are extracted from the frequency domain signals, and the power factor and total harmonic distortion are calculated.

[0062] S2. Feature image processing, converting the frequency domain features into a two-dimensional image through discrete cosine transform (DCT), the two-dimensional image includes the feature information of the current and the feature information of the power, each feature value in the two-dimensional image corresponds to a pixel value in the image, thereby converting the one-dimensional feature vector into a two-dimensional image;

[0063] S3. Feature image fusion, using Laplace pyramid fusion technology, merge the current feature image and the power feature image into an RGB image. In the process, the current feature image is used as the red channel and the power feature image is used as the green channel. If necessary, other channels can be added.

[0064] S4. Model training: Select a lightweight deep learning model (EfficientNet) and use the fused feature images for training. By adjusting the weights and biases of the model, the model can learn to recognize different load feature images.

[0065] The S1 includes:

[0066] S11. Data acquisition: At the power supply inlet, high-precision, low-temperature-coefficient constantan resistors are used, and their resistance values ​​are strictly calibrated to ensure that the measurement error under different currents is extremely small. A professional-grade high-speed data acquisition card is used, and its sampling frequency is set between 10kHz and 20kHz, which can be flexibly adjusted according to the actual load characteristics. Taking 12kHz as an example, the acquisition card synchronously and accurately acquires the current and voltage in the load circuit, and the acquisition time lasts from several seconds to several minutes, ensuring that various operating states of the load are covered and sufficient raw data samples are obtained.

[0067] S12. Frequency domain conversion and harmonic analysis: Fast Fourier transform is performed on the collected time domain signal. For discrete time signals, the discrete Fourier transform (DFT) formula is:

[0068]

[0069] Where n is the index of the sequence; k is the frequency domain index, which indicates the position of different frequency components in the frequency domain; x(n) is the discrete time signal; N is the length of the discrete sequence; is a complex exponential function, where j is the imaginary unit (j 2 =-1), the complex exponential function plays a key role in mapping the time domain signal to the frequency domain in DFT. Its periodicity and orthogonality make it possible to decompose the time domain signal into a combination of sine and cosine components of different frequencies through this formula;

[0070] The frequency domain signal is calculated by the FFT algorithm, and the amplitude and phase of the fundamental frequency and each harmonic (usually considering the 2nd to 50th harmonic) are extracted;

[0071] During the extraction process, a peak detection algorithm and a phase calculation function are used to accurately locate the harmonic frequency components and their corresponding amplitude and phase information;

[0072] S13. Power factor and total harmonic distortion calculation,

[0073] First, calculate the active power. The active power is calculated using the numerical integration method. The formula is as follows:

[0074]

[0075] Where T is the period of the current and voltage signals, which reflects the repeatability of the signals in time; M is the number of time intervals after discretization, and Δt is the time interval; u(t) and i(t) are the voltage and current signals that change with time, respectively, and m is the index of the discrete time point; u(mΔt) and i(mΔt) are the voltage and current values ​​at the discrete time point mΔt;

[0076] Then, calculate the apparent power. When calculating the apparent power, first calculate the effective values ​​of voltage and current respectively through the root mean square algorithm.

[0077] Voltage effective value U:

[0078]

[0079] Current effective value I:

[0080]

[0081] Apparent power:

[0082] S = UI;

[0083] Finally, calculate the total harmonic distortion using the formula:

[0084]

[0085] Among them, A n It represents the amplitude of the nth harmonic, A1 is the amplitude of the fundamental wave; the fundamental wave is the main frequency component of the signal, and the harmonics are the frequency components other than the fundamental wave.

[0086] In S2, the frequency domain features extracted by S1 are input into a two-dimensional discrete cosine transform (DCT) module. For a two-dimensional signal, the DCT transform formula is:

[0087]

[0088] x(n,m) is the frequency domain feature data obtained by S1; C(u,v) is the result of two-dimensional discrete cosine transform (DCT); u and v are the coordinates in the two-dimensional image; α(v) and u(v) are coefficients;

[0089] Through the above transformation, the one-dimensional frequency domain feature vector is converted into a two-dimensional image, in which each pixel value corresponds to a DCT coefficient, and the row and column structure of the image intuitively presents the spatial distribution characteristics of the frequency domain features, thereby generating a two-dimensional image containing the characteristic information of the current (such as the distribution characteristics of the current harmonics in the frequency domain) and the characteristic information of the power (such as the power factor, the manifestation of the total harmonic distortion in the frequency domain, etc.).

[0090] In S3, the step of feature image fusion includes:

[0091] S31. Laplacian pyramid construction, constructing Laplacian pyramids for the current feature image and the power feature image respectively;

[0092] Taking the current characteristic image as an example, we first construct a Gaussian pyramid and set the original image as the 0th layer, that is, Then, after Gaussian blurring the current layer image (using a 5x5 Gaussian kernel, for example), alternate row and column sampling is performed to obtain the next layer image, that is, Repeat this process to construct L-layer Gaussian pyramid (L is determined according to the image size and fusion requirements, such as 3-5 layers); then calculate the Laplacian pyramid: Expand is an upsampling operation, which starts from the L-1th layer and calculates upward to obtain the Laplacian pyramid representation of the current feature image:

[0093]

[0094] The same method is used to construct the Laplacian pyramid of the power feature image;

[0095] S32. Image fusion: The Laplacian pyramid of the current feature image is used as the red channel of the RGB image; the Laplacian pyramid of the power feature image is used as the green channel. If other feature images (such as image representations of some special load characteristics found in subsequent studies) need to be added, their Laplacian pyramids can be constructed according to the same process and assigned to corresponding channels (such as the blue channel), thereby realizing multi-feature image fusion and generating the Laplacian pyramid representation of the fused image:

[0096]

[0097] S33. Reconstruct the fused image. Reconstruct the image from the lowest level of the fused Laplacian pyramid. First, Upsample to get the upsampled image Then with Add, that is,

[0098] Then The upsampling and addition operations are repeated, and so on, until the final reconstructed RGB fused image is obtained, which completely retains the fused multi-feature information and provides rich input data for subsequent model training.

[0099] The S4 comprises the following steps:

[0100] S41. Data preparation and enhancement: The fused feature image dataset is divided into training set, validation set and test set in a ratio of 7:2:1. Multi-dimensional enhancement operations are performed on the training set data, including random rotation (angle range is [-15°, 15°], using bilinear interpolation to maintain image quality), horizontal and vertical flipping, random cropping (randomly cropping 5%-10% at the edge of the image, and the image size is adjusted by padding or scaling after cropping), and adding Gaussian noise (mean is 0, standard deviation is between 0.01-0.05, dynamically adjusted according to the image pixel value range), etc., which greatly expands the diversity of training data and improves the generalization ability of the model;

[0101] S42. Model selection and initialization. According to the complexity of load characteristics and the size of data sets, for relatively simple load scenarios and small data sets, select the appropriate version in the lightweight deep learning model EfficientNet family, and select EfficientNet-B0; for complex loads and large data sets, you can choose EfficientNet-B4, etc.; use the Kaiming initialization method to initialize the weights of the convolution layer, calculate the appropriate initialization weight variance according to the number of input and output channels of the convolution layer, and initialize the bias to 0. At the same time, set the initial learning rate between 0.001-0.01 (such as 0.005), the weight decay coefficient between 0.0001-0.001 (such as 0.0005), the momentum parameter to 0.9 and other hyperparameters to build the initial environment for model training;

[0102] S43. Training and optimization, defining the cross entropy loss function Among them, C is the number of categories; y is the true label; To predict the output, the Adam optimizer is selected, and its parameter update formula is:

[0103] m t =β1m t-1 +(1-β1)g t

[0104]

[0105] Among them, β1 = 0.9; β2 = 0.999; η is the learning rate; ε = 1e-8; g t is the current gradient, θ t are model parameters; during the training process, image data is input into the model in batches of 32-128 (such as 64), and forward propagation is performed to calculate the predicted output. After calculating the loss function, the model parameters are updated through back propagation; after each training epoch, the model performance is evaluated on the validation set, and indicators such as accuracy, recall rate, and F1 value are calculated. According to the results of the validation set, the hyperparameters (such as the learning rate attenuation factor and whether to stop training early) are dynamically adjusted to ensure the effectiveness and stability of the model training.

[0106] An epoch represents a complete training iteration for the entire training dataset; that is, an epoch is completed when the model performs a forward propagation and a backpropagation operation using each sample in the training dataset.

[0107] The above description is a preferred embodiment of the invention and is not intended to limit the invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the invention should be included in the protection scope of the invention.

Claims

1. A non-intrusive load identification method based on high-frequency feature imaging of loads, characterized in that: The following steps are involved: S1. Data collection and feature extraction. At the power supply inlet, the current and voltage data in the load circuit are collected by connecting a constantan resistor in series. The collected time domain signals are converted into frequency domain signals using Fourier transform. The fundamental frequency and the amplitude and phase of each harmonic are extracted from the frequency domain signals, and the power factor and total harmonic distortion are calculated. S2. Feature image processing, converting the frequency domain features into a two-dimensional image through discrete cosine transform (DCT), the two-dimensional image includes the feature information of the current and the feature information of the power, each feature value in the two-dimensional image corresponds to a pixel value in the image, thereby converting the one-dimensional feature vector into a two-dimensional image; S3. Feature image fusion, using Laplace pyramid fusion technology, merge the current feature image and the power feature image into an RGB image. In the process, the current feature image is used as the red channel and the power feature image is used as the green channel; S4. Model training: Select a lightweight deep learning model (EfficientNet) and use the fused feature images for training. By adjusting the weights and biases of the model, the model can learn to recognize different load feature images.

2. The non-intrusive load identification method based on high-frequency feature imaging of load according to claim 1 is characterized in that: The S1 includes: S11. Data acquisition: At the power supply inlet, high-precision, low-temperature-coefficient constantan resistors are used, and their resistance values ​​are strictly calibrated to ensure that the measurement error under different currents is minimal; using a data acquisition card, the sampling frequency is set between 10kHz and 20kHz, which can be flexibly adjusted according to the actual load characteristics; S12. Frequency domain conversion and harmonic analysis: Fast Fourier transform is performed on the collected time domain signal. For discrete time signals, the discrete Fourier transform (DFT) formula is: Where n is the index of the sequence; k is the frequency domain index, which indicates the position of different frequency components in the frequency domain; x(n) is the discrete time signal; N is the length of the discrete sequence; is a complex exponential function, where j is the imaginary unit (j 2 =-1), the complex exponential function plays a key role in mapping the time domain signal to the frequency domain in DFT; The frequency domain signal is calculated by FFT algorithm, and the fundamental frequency and the amplitude and phase of each harmonic are extracted; During the extraction process, a peak detection algorithm and a phase calculation function are used to accurately locate the harmonic frequency components and their corresponding amplitude and phase information; S13. Power factor and total harmonic distortion calculation, First, calculate the active power. The active power is calculated using the numerical integration method. The formula is as follows: Where T is the period of the current and voltage signals, which reflects the repeatability of the signals in time; M is the number of time intervals after discretization, and Δt is the time interval; u(t) and i(t) are the voltage and current signals that change with time, respectively, and m is the index of the discrete time point; u(mΔt) and i(mΔt) are the voltage and current values ​​at the discrete time point mΔt; Then, calculate the apparent power. When calculating the apparent power, first calculate the effective values ​​of voltage and current respectively through the root mean square algorithm. Voltage effective value U: Current effective value I: Apparent power: S = UI; Finally, calculate the total harmonic distortion using the formula: Among them, An represents the amplitude of the nth harmonic, and A1 is the amplitude of the fundamental wave; the fundamental wave is the main frequency component of the signal, and the harmonics are other frequency components except the fundamental wave.

3. The non-intrusive load identification method based on high-frequency feature imaging of load according to claim 1 is characterized in that: In S2, the frequency domain features extracted by S1 are input into a two-dimensional discrete cosine transform (DCT) module. For a two-dimensional signal, the DCT transform formula is: x(n,m) is the frequency domain feature data obtained by S1; C(u,v) is the result of the two-dimensional discrete cosine transform (DCT); u and v are the coordinates in the two-dimensional image; α(v) and u(v) are coefficients; Through the above transformation, the one-dimensional frequency domain feature vector is converted into a two-dimensional image, in which each pixel value corresponds to a DCT coefficient, and the row and column structure of the image intuitively presents the spatial distribution characteristics of the frequency domain features, thereby generating a two-dimensional image containing the characteristic information of the current and the characteristic information of the power.

4. The non-intrusive load identification method based on high-frequency feature imaging of load according to claim 1 is characterized in that: In S3, the step of feature image fusion includes: S31. Laplacian pyramid construction, constructing Laplacian pyramids for the current feature image and the power feature image respectively; Taking the current characteristic image as an example, we first construct a Gaussian pyramid and set the original image as the 0th layer, that is, Then, after Gaussian blurring the current layer image, alternate row and column sampling is performed to obtain the next layer image, that is, Repeat this process to construct L-layer Gaussian pyramid (L is determined according to the image size and fusion requirements, such as 3-5 layers); then calculate the Laplacian pyramid: Expand is an upsampling operation, which starts from the L-1th layer and calculates upward to obtain the Laplacian pyramid representation of the current feature image: The same method is used to construct the Laplacian pyramid of the power feature image; S32. Image fusion: The Laplacian pyramid of the current feature image is used as the red channel of the RGB image; the Laplacian pyramid of the power feature image is used as the green channel. If other feature images need to be added, their Laplacian pyramids can be constructed according to the same process and assigned to the corresponding channels, thereby realizing multi-feature image fusion and generating the Laplacian pyramid representation of the fused image: S33. Reconstruct the fused image. Reconstruct the image from the lowest level of the fused Laplacian pyramid. First, Upsample to get the upsampled image Then with Add, that is, Then The upsampling and addition operations are repeated, and so on, until the final reconstructed RGB fused image is obtained.

5. The non-intrusive load identification method based on high-frequency feature imaging of load according to claim 1 is characterized in that: The S4 comprises the following steps: S41. Data preparation and enhancement: The fused feature image dataset is divided into training set, validation set and test set in a ratio of 7:2:

1. Multi-dimensional enhancement operations are performed on the training set data, including random rotation, horizontal and vertical flipping, random cropping and adding Gaussian noise, which greatly expands the diversity of training data and improves the generalization ability of the model. S42. Model selection and initialization. According to the complexity of load characteristics and the size of data sets, for relatively simple load scenarios and small data sets, select the appropriate version in the lightweight deep learning model EfficientNet family, and select EfficientNet-B0; for complex loads and large data sets, EfficientNet-B4 can be selected; Kaiming initialization method is used to initialize the weights of the convolution layer, and the appropriate initialization weight variance is calculated according to the number of input and output channels of the convolution layer. The bias is initialized to 0. At the same time, the initial learning rate is set between 0.001-0.01, the weight decay coefficient is set between 0.0001-0.001, and the momentum parameter is set to 0.9 hyperparameters to build the initial environment for model training; S43. Training and optimization, defining the cross entropy loss function Among them, C is the number of categories; y is the true label; To predict the output, the Adam optimizer is selected, and its parameter update formula is: m t =β1m t-1 +(1-β1)g t ; Among them, β1 = 0.9; β2 = 0.999; η is the learning rate; ε = 1e-8; g t is the current gradient, θ t are model parameters; during the training process, image data is input into the model in batches of 32-128, and forward propagation is performed to calculate the predicted output. After calculating the loss function, the model parameters are updated through back propagation; after each training epoch, the model performance is evaluated on the validation set, and the accuracy, recall rate, and F1 value indicators are calculated. The hyperparameters are dynamically adjusted according to the validation set results to ensure the effectiveness and stability of the model training.

Images