A non-intrusive load identification method based on multi-dimensional feature fusion image

Abstract

This invention discloses a non-intrusive load identification method based on multi-dimensional feature fusion images. The method includes data preprocessing to obtain an S-mode matrix, a power factor-weighted voltage-reactive power trajectory, and a power-weighted recurrence graph matrix; feature fusion image generation, where the three matrices are used as R, G, and B channels respectively to obtain a feature-fused color image; and load identification, where the generated color image is input into a ResNet-18 model pre-trained on ImageNet for transfer learning to complete load identification. This invention utilizes a combined feature fusion approach, and finally inputs the generated color image into a ResNet-18 model pre-trained on ImageNet for transfer learning to complete load identification. This helps prevent load identification from occurring when identifying appliances with multiple states and also improves the accuracy of identifying resistive loads.

Description

Technical Field

[0001] This invention relates to the field of non-invasive load recognition, and in particular to a non-invasive load recognition method based on multi-dimensional feature fusion images. Background Technology

[0002] With the development of the economy and society, the demand for electricity is constantly increasing, and it is urgent to manage electricity consumption. Understanding residential load information is of great significance for load management, which necessitates the use of non-intrusive load monitoring technology.

[0003] Existing non-intrusive load monitoring technology can analyze the types, operating status and energy consumption information of residential loads by collecting power consumption data at the bus. However, this monitoring method has limited image features, which can lead to confusion in the identification of appliances with multiple states, and can easily cause misjudgment in the identification of resistive loads. Summary of the Invention

[0004] The purpose of this invention is to provide a non-invasive load recognition method based on multi-dimensional feature fusion images to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a non-invasive load recognition method based on multi-dimensional feature fusion images, comprising the following specific steps:

[0006] Step 1: Data preprocessing. The high-frequency acquired current and voltage data are processed to obtain the S-mode matrix, the power factor-weighted voltage-reactive power trajectory, and the power-weighted recursive graph matrix.

[0007] Step 2: Feature fusion image generation. The S-mode matrix, the power factor-weighted voltage-reactive power trajectory, and the power-weighted recursive graph matrix are used as the R, G, and B channels, respectively, to obtain a feature-fused color image.

[0008] Step 3: Load identification. The generated color image is input into a ResNet-18 model pre-trained on ImageNet for transfer learning to complete load identification.

[0009] Preferably, the S-modulus matrix in step one is obtained by performing an S-transform on the various load current data acquired at high frequency under steady state. After the S-transform, the current data yields a complex matrix, and the S-modulus matrix is ​​obtained by taking its modulus. The S-transform is used to reveal the time-domain and frequency-domain characteristics of the signal. The S-transform can be used to localize the signal in the time domain through a window function, and then a Fourier transform is performed to obtain the representation in the time-frequency domain. For the input signal, the S-transform is defined as follows:

[0010]

[0011]

[0012] Where w(t-τ,f) represents the Gaussian window function, τ is the translation factor that controls the position of the Gaussian window on the time axis, f is the frequency, and the window width factor is σ = 1 / f;

[0013] The S-transform yields a two-dimensional complex matrix. The S-modulus matrix, obtained by modulo the two-dimensional complex matrix, is then used to select the load characteristic harmonics as the R-channel. The formula for the S-modulus matrix is:

[0014] X(τ,f)=|S(τ,f)|

[0015] Where X represents the S-modulus matrix obtained by taking the modulus of a two-dimensional complex matrix.

[0016] Preferably, in step one, the power factor-weighted voltage-reactive power trajectory is differentiated by acquiring current data. This involves extracting one cycle of current and voltage data from the voltage zero-crossing point under steady-state conditions, and then, based on Fryze power theory, decomposing the load current into active and reactive current. The formula for calculating the reactive current is:

[0017] i f (t)=i(t)-i a (t)

[0018] Among them, i f (t) reactive component of load current, i(t) original load current, i a (t) represents the active component of the load current, i. a The formula for calculating (t) is:

[0019]

[0020]

[0021] Where P represents the active power of the load, N represents the number of data points in one cycle, and i d and u d These are the d-th current and voltage values ​​within the period, U rms This is the effective value of the voltage;

[0022] After obtaining the voltage and reactive current data for the next cycle in steady state, they are normalized, and the calculation formula is as follows:

[0023]

[0024]

[0025] Among them, i minand i max These represent the minimum and maximum values ​​of the reactive current during the period, u and u, respectively. min and u max These represent the minimum and maximum voltage values ​​during the period, respectively.

[0026] Create a VI using normalized data. f The trajectory is covered using an M×M zero matrix. Then, all sampling points are traversed. If a sampling point is within the matrix range, the matrix cell value is set to 1. Finally, a representative VI is constructed. f The grayscale matrix of the trajectory;

[0027] Power factor is an important characteristic of a load, representing the load's efficiency in utilizing the active power of the power source. The power factor is calculated as follows:

[0028] S rms =I rms *U rms

[0029] λ = P / S rms

[0030] Among them, I rms S is the effective value of the current. rms VI represents apparent power, λ represents the load power factor, and VI represents apparent power. f Multiplying the gray value matrix of the trajectory by λ yields the power factor-weighted VI. f Trajectory matrix.

[0031] Preferably, the recursive graph of the power-weighted recursive graph matrix in step one can reveal the internal structure of the time series. The recursive graph matrix is ​​obtained by transforming the time domain space of the time series to the phase space, calculating the distance between every two states, and finally performing threshold binarization. The recursive formula can be represented by the recursive matrix R. ij The formula is as follows:

[0032] R ij =θ(ε-E) ij ), i,j=1,…N

[0033] E ij =||X i -X j ||

[0034] Among them, E ij Represents vector X i and X j The Euclidean distance between them, ε is the distance threshold, and θ(·) represents the Heaviside function, expressed as:

[0035]

[0036] Load power is an important characteristic of the load, but this information is completely lost after normalization. Therefore, the power information is weighted onto the recursive matrix, and the calculation method is defined as follows:

[0037]

[0038] R0=σR ij

[0039] Among them, S rms S represents the apparent power of the load. max This represents the maximum apparent power of all loads, where the power of some appliances is related to S. max If the difference is too large, the coefficient σ is squared, and R0 represents the final power-weighted recursive matrix.

[0040] Preferably, the feature fusion color image in step two refers to merging different feature information to form a color image with comprehensive features. The feature fusion method is weighted summation, local region merging, and multi-scale fusion. The feature fusion color image can provide a more comprehensive and accurate image description, providing more information for image analysis and application.

[0041] Preferably, the ImageNet dataset in step three is a large-scale image database used for computer vision research and algorithm evaluation. The ImageNet dataset contains millions of images from various fields, including a variety of objects, scenes, and concepts.

[0042] Preferably, the ResNet-18 model in step three is a deep residual network architecture that adopts the idea of ​​residual learning. It solves the degradation in deep neural network training by introducing residual connections. The residual connections allow the network to skip a layer during information flow, enabling the network to learn features better.

[0043] Preferably, the transfer learning in step three can utilize the already learned knowledge and experience to improve learning performance on a new task.

[0044] The technical effects and advantages of this invention are as follows:

[0045] This invention utilizes a feature fusion approach, processing high-frequency acquired current and voltage data to obtain an S-mode matrix, a power factor-weighted voltage-reactive power trajectory, and a power-weighted recurrence graph matrix. These three matrices are then fused into a feature-rich color image. Finally, the generated color image is input into a ResNet-18 model pre-trained on ImageNet for transfer learning to complete load identification. The accuracy rates achieved on the public datasets PLAID and WHITED are 99.11% and 99.69%, respectively, and on the self-test dataset, a recognition accuracy of 99.24%. This helps prevent confusion when identifying appliances with multiple states and also improves the accuracy of identifying resistive loads, preventing misjudgments. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the non-invasive load identification process of the present invention.

[0047] Figure 2 This is a schematic diagram of the multidimensional feature image synthesis process of the present invention.

[0048] Figure 3 This is a color image schematic diagram of eleven electrical appliances of the present invention.

[0049] Figure 4 This is a schematic diagram illustrating the training results of transfer learning and non-transfer learning in this invention.

[0050] Figure 5 This is a schematic diagram of the confusion matrix of the experimental results of the PLAID test set of this invention.

[0051] Figure 6 This is a schematic diagram of the confusion matrix of the experimental results of the WHITED test set of this invention.

[0052] Figure 7 This is a schematic diagram of the color feature images of a portion of the resistive load in the WHITED dataset of this invention.

[0053] Figure 8 This is a schematic diagram of the confusion matrix of the experimental results of the self-test test set of this invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] This invention provides, for example Figure 1-8This paper presents a non-intrusive load identification method based on multi-dimensional feature fusion images. The method includes data preprocessing, which processes high-frequency acquired current and voltage data to obtain an S-mode matrix, a power factor-weighted voltage-reactive power trajectory, and a power-weighted recursive graph matrix. Feature fusion image generation is then performed. The feature fusion color image refers to the merging of different feature information to form a color image with comprehensive features. The feature fusion methods include weighted summation, local region merging, and multi-scale fusion. The feature-fused color image can provide a more comprehensive and accurate image description, offering more information for image analysis and applications. This fusion method provides richer and more accurate image information, helping people to gain a deeper understanding and analysis of images. Analysis shows that common features in feature-fused color images include color, texture, and shape. These features can be extracted using different image processing and analysis methods and then fused. The S-mode matrix, the power factor-weighted voltage-reactive power trajectory, and the power-weighted recursive graph matrix are used as the R, G, and B channels, respectively, to obtain a feature-fused color image. For load identification, the generated color image is input into a ResNet-18 model pre-trained on ImageNet for transfer learning to complete load identification. The ImageNet dataset is a large-scale image database used for computer vision research and algorithm evaluation, containing millions of images from various fields, including various objects. The ImageNet dataset, divided into training, validation, and test sets, aims to advance image classification to a higher level, enabling computers to recognize and understand more complex visual information. The ResNet-18 model is a deep residual network architecture that employs residual learning. It addresses the degradation in deep neural network training by introducing residual connections. Residual connections allow the network to skip one or more layers during information flow, enabling it to learn features more effectively. ResNet-18 contains 18 convolutional and fully connected layers, including four large convolutional blocks, each consisting of two convolutional layers and a residual connection. Batch normalization and activation functions were used for feature extraction and nonlinear activation. Finally, global average pooling and a fully connected layer containing the number of classes were used for classification. Transfer learning is a machine learning technique that leverages learned knowledge and experience to improve learning performance on a new task. It involves using the parameters or features of one or more pre-trained models to solve a related target task. The main idea of ​​transfer learning is that knowledge from the source task can aid in learning the target task, even if the two tasks have different data distributions or feature representations. Non-invasive load identification refers to identifying the operating status and load characteristics of load equipment through monitoring and analysis of power grid signals without electrical wiring or sensor installation.This identification method is typically based on signal processing technology and machine learning algorithms. Non-intrusive load identification enables real-time monitoring and analysis of different types of load devices in the power grid, including load changes, power consumption, and usage patterns. Through time-domain, frequency-domain, and waveform analysis of power grid signals, the electrical characteristics of load devices can be obtained, and load type and status can be identified accordingly. Non-intrusive load identification has significant application value in smart grids, energy management, and power system optimization. It helps power system operators understand users' electricity consumption in real time, optimize grid dispatching and load management, and improve energy efficiency and power supply quality. Simultaneously, non-intrusive load identification also provides residents and businesses with tools for electricity monitoring and management, helping them achieve energy conservation, emission reduction, and electricity cost control goals.

[0056] The S-modulus matrix is ​​obtained by performing an S-transform on the high-frequency acquired load current data under steady-state conditions. The current data, after the S-transform, yields a complex matrix. Taking the modulus of this matrix gives the S-modulus matrix. The S-transform reveals the time-domain and frequency-domain characteristics of a signal. The S-transform can be used to localize the signal in the time domain using a window function, followed by a Fourier transform to obtain its representation in the time-frequency domain. For the input signal, the S-transform is defined as follows:

[0057]

[0058]

[0059] Where w(t-τ,f) represents the Gaussian window function, τ is the translation factor that controls the position of the Gaussian window on the time axis, f is the frequency, and the window width factor is σ = 1 / f;

[0060] The S-transform yields a two-dimensional complex matrix. The S-modulus matrix, obtained by modulo the two-dimensional complex matrix, is then used to select the load characteristic harmonics as the R-channel. The formula for the S-modulus matrix is:

[0061] X(τ,f)=|S(τ,f)|

[0062] Where X represents the S-modulus matrix obtained by taking the modulus of a two-dimensional complex matrix.

[0063] The power factor-weighted voltage-reactive power trajectory is differentiated by acquiring current data. This involves capturing one cycle of current and voltage data from the voltage rise-to-zero point under steady-state conditions. Then, based on Fryze power theory, the load current is decomposed into active and reactive currents. The formula for calculating the reactive current is:

[0064] i f (t)=i(t)-i a (t)

[0065] Among them, if (t) reactive component of load current, i(t) original load current, i a (t) represents the active component of the load current, i. a The formula for calculating (t) is:

[0066]

[0067]

[0068] Where P represents the active power of the load, N represents the number of data points in one cycle, and i d and u d These are the d-th current and voltage values ​​within the period, U rms This is the effective value of the voltage;

[0069] After obtaining the voltage and reactive current data for the next cycle in steady state, they are normalized, and the calculation formula is as follows:

[0070]

[0071]

[0072] Among them, i min and i max These represent the minimum and maximum values ​​of the reactive current during the period, u and u, respectively. min and u max These represent the minimum and maximum voltage values ​​during the period, respectively.

[0073] Create a VI using normalized data. f The trajectory is covered using an M×M zero matrix. Then, all sampling points are traversed. If a sampling point is within the matrix range, the matrix cell value is set to 1. Finally, a representative VI is constructed. f The grayscale matrix of the trajectory;

[0074] Power factor is an important characteristic of a load, representing the load's efficiency in utilizing the active power of the power source. The power factor is calculated as follows:

[0075] S rms =I rms *U rms

[0076] λ = P / S rms

[0077] Among them, I rms S is the effective value of the current. rms VI represents apparent power, λ represents the load power factor, and VI represents apparent power. f Multiplying the gray value matrix of the trajectory by λ yields the power factor-weighted VI.f Trajectory matrix.

[0078] Power-weighted recursive graph matrices are a commonly used processing and analysis method in the study of non-stationary signals. They can reveal the internal structure of time series. The recursive graph matrix transforms the time series from the time domain to the phase space, calculates the distance between every two states, and finally performs threshold binarization to obtain the result. The recursive expression can be represented by the recursive matrix R. ij The formula is as follows:

[0079] R ij =θ(ε-E) ij ), i,j=1,…N

[0080] E ij =||X i -X j ||

[0081] Among them, E ij Represents vector X i and X j The Euclidean distance between them, ε is the distance threshold, and θ(·) represents the Heaviside function, expressed as:

[0082]

[0083] Load power is an important characteristic of the load, but this information is completely lost after normalization. Therefore, the power information is weighted onto the recursive matrix, and the calculation method is defined as follows:

[0084]

[0085] R0=σR ij

[0086] Among them, S rms S represents the apparent power of the load. max This represents the maximum apparent power of all loads, where the power of some appliances is related to S. max If the difference is too large, the coefficient σ is squared, and R0 represents the final power-weighted recursive matrix.

[0087] This invention is achieved through the following process, combined with Figure 1 As shown, the process mainly includes three steps: data preprocessing, feature fusion image generation, and load recognition.

[0088] Data preprocessing begins by performing an S-transform on the high-frequency acquired load current data under steady-state conditions. The resulting complex matrix is ​​then moduloed to obtain the S-modulus matrix. The load characteristic harmonics are selected as the R-channel of the image. Next, current and voltage data for one steady-state cycle are extracted, starting from the voltage zero-crossing point. The reactive current of the load is obtained using Fryze power theory, and a binarized VI plot is then generated. f The trajectory is then used to calculate the load power factor based on the load's active power and apparent power, and weighted in VI. f The trajectory is used as the G channel. The recursive graph analysis is performed again on the periodic current sequence collected in the previous step to obtain the recursive graph matrix. Then, the load power information is weighted into the recursive graph matrix as the B channel. Finally, the three matrices obtained in the first three steps are scaled down to the same size and fused into a three-channel color image with rich feature information.

[0089] Feature fusion image generation, taking the 10th CSV file of the PLAID dataset as an example, the grayscale images of the three matrices (M=64) generated according to the above method are attached. Figure 2 As shown in (a), (b), and (c), these are used as the three-channel inputs of the RGB image to synthesize a color image with feature fusion, as follows. Figure 2 As shown in (d), the color maps of load characteristics generated for 11 types of electrical appliances in the PLAID dataset are attached. Figure 3 As shown, the air conditioner and refrigerator have multiple states;

[0090] Load identification begins with the experimental environment and evaluation metrics. To better verify the effectiveness of the algorithm of this invention, the environmental parameters for this embodiment are shown in Table 1, and the specific experimental settings are as follows:

[0091]

[0092] Table 1

[0093] This invention uses accuracy, confusion matrix, and F1 score as evaluation metrics for load recognition. Accuracy is used to evaluate the overall recognition performance of the dataset, and is calculated as follows:

[0094]

[0095] In the formula, A represents the total number of samples, and a represents the number of correctly identified samples;

[0096] The F1 score is used to evaluate the recognition performance for each type of load, and the calculation formula is as follows:

[0097]

[0098]

[0099]

[0100] In the formula, TP represents the number of samples that are actually positive and are identified as positive, FP represents the number of samples that are actually negative but are identified as positive, FN represents the number of samples that are actually positive but are identified as negative, Pre is the precision, and Re is the recall.

[0101] Secondly, the main experimental results include the results of the PLAID dataset, the WHITED dataset, the comparison of results of different visualization methods, and the results of the self-test dataset.

[0102] Experimental results on the PLAID dataset: The PLAID dataset contains 1074 instances of 15 electrical appliances. Detailed information from startup to stable operation was recorded at a sampling frequency of 30kHz. Based on this invention, the voltage and current in the PLAID dataset were processed. A load image set with fused features was generated every 20 cycles, totaling 5066 samples. These samples were divided into training, validation, and test sets in a 6:2:2 ratio. Training used the Adam optimizer with an adaptive learning rate adjustment mechanism. The initial learning rate was 0.0001. If the loss function did not decrease within 7 iterations, the learning rate was updated to 1 / 5 of the original learning rate to optimize the model. The batch size for model training was 64, and training consisted of 70 iterations. Figure 4 The figures show the changes in accuracy and loss function of the PLAID validation set with increasing training iterations under both transfer learning and non-transfer learning conditions. It can be seen that training based on transfer learning converges faster and with less fluctuation. Furthermore, the average training time per iteration is 21.4 seconds without transfer learning, while it only takes 18.5 seconds with transfer learning, demonstrating the superiority of transfer learning. The validation set accuracy reaches 95% at the 20th training iteration, and complete convergence occurs around the 60th iteration, ultimately reaching 99.51% accuracy on the validation set. When the test set is input into the trained ResNet-18 model, the final test set accuracy reaches 99.11%, with an average F1 score of 98.97%. This shows that the method of this invention has good recognition performance. The confusion matrix results are attached. Figure 5 As shown in the diagram, by observing the confusion matrix, it can be seen that refrigerators and air conditioners are prone to misjudgment, with large differences in characteristics under different states, making them easy to be misjudged with other appliances. The harmonic information of fluorescent lamps and laptops is relatively similar, making them easy to confuse. Heaters and hair dryers are both heating devices with similar internal components, and they are also confused. Microwave ovens, vacuum cleaners, and washing machines were not misjudged. Among them, microwave ovens produce a larger third harmonic, which makes them more distinguishable from other appliances. Washing machines and vacuum cleaners have the most stable operating states and are not easily confused.

[0103] Experimental results for the WHITED dataset: The WHITED dataset contains 54 types of appliances, with a total of 1339 instances and a sampling frequency of 44.1kHz. Similar experimental procedures to the PLAID dataset were used for the WHITED dataset, employing a resolution of 64, generating a total of 9592 samples. The confusion matrix for the test set is attached. Figure 6 As shown, the final accuracy reached 99.68%, demonstrating equally excellent recognition performance on the WHITED dataset, with only irons and kettles being confused. The average F1 score reached 99.88%. The WHITED dataset contains many resistive heating appliances; the results show that only irons and kettles, both belonging to the high-power resistive load category, were misclassified because their harmonics are almost zero. f The trajectories are similar, and the power is all above 1600W. Other resistance heating appliances are correctly identified. This is because channel B of this invention is based on a power-weighted recursive graph, which easily identifies appliances with different power ratings. (See attached image.) Figure 7 These are characteristic images of three resistive loads: a white kettle, a water heater, and a toaster. The images include harmonic information and VI. f The trajectory is difficult to identify. The power level determines the depth value of channel B, so resistive loads with different power can be correctly identified.

[0104] Comparison of Results from Different Image Recognition Methods: To fully verify the effectiveness of the method of this invention, the load recognition performance of this invention was compared with that of current state-of-the-art methods on the PLAID and WHITED datasets. The recognition accuracy of different methods is shown in Table 2. The results show that the accuracy of the method of this invention is better than the other four load recognition methods. Method 1 uses weighted grayscale value VI as the load feature, and Method 2 uses weighted recursive graph as the load. The feature extraction of these two methods is too simple, so the accuracy is not high. Compared with Method 3, the accuracy of the method of this invention is improved by 1.72% and 1.10% respectively on the two datasets. f Color coding of the trajectory results in a large number of blank areas, making it difficult for neural networks to capture some feature information on the trajectory. Method 4 reduces the load VI. f The trajectory, current trajectory, and instantaneous power trajectory are combined into a single true-color image as the load feature. However, resistive loads have three similar trajectories, which can easily lead to confusion. Through comparative experiments, the method of this invention achieved higher recognition accuracy on both datasets. The recognition accuracy of different methods is shown in Table 2 below:

[0105]

[0106] Table 2

[0107] Experimental results of the self-test dataset: The self-test dataset contains 12 types of electrical appliances from 10 households, two dormitory buildings, and two laboratories, totaling 719 instances. The types and quantities of electrical appliances are shown in Table 3.

[0108]

[0109]

[0110] Table 3

[0111] The 719 sets of self-test data were preprocessed to generate a total of 7205 samples. The confusion matrix of the load identification test set using the method of this invention is shown in the appendix. Figure 8 As shown, due to the large power variation of variable frequency air conditioners and the fact that they include both heating and cooling states, other electrical appliances are easily misidentified as air conditioners. The overall recognition accuracy rate reaches 99.24%, and the average F1 score reaches 99.25%. It can be seen that the method of the present invention also has a high recognition accuracy rate on the self-test dataset.

[0112] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-invasive load recognition method based on multi-dimensional feature fusion images, characterized in that, The specific steps include the following: Step 1: Data preprocessing. The high-frequency acquired current and voltage data are processed to obtain the S-mode matrix, the power factor-weighted voltage-reactive power trajectory, and the power-weighted recursion graph matrix. The recursion graph matrix reveals the internal structure of the time series. The recursion graph matrix is ​​obtained by transforming the time-domain space of the time series to the phase space, calculating the distance between every two states, and finally performing threshold binarization. The recursive expression is represented by the recursion matrix. The formula is as follows: in, Representing vectors and The Euclidean distance between them Distance threshold The Heaviside function is represented as: Load power is an important characteristic of the load, but this information is completely lost after normalization. Therefore, the power information is weighted onto the recursive matrix, and the calculation method is defined as follows: in, This represents the apparent power of the load. This represents the maximum apparent power of all loads, where the power of some appliances is related to... The difference is too large, which in turn affects the coefficient. Take the square root. This represents the final power-weighted recursive matrix; Step 2: Feature fusion image generation. The S-mode matrix, the power factor-weighted voltage-reactive power trajectory, and the power-weighted recursive graph matrix are used as the R, G, and B channels, respectively, to obtain a feature-fused color image. Step 3: Load identification. The generated color image is input into a ResNet-18 model pre-trained on ImageNet for transfer learning to complete load identification.

2. The non-invasive load recognition method based on multi-dimensional feature fusion image according to claim 1, characterized in that, The S-modulus matrix in step one is obtained by performing an S-transform on the high-frequency acquired load current data under steady-state conditions. The current data, after the S-transform, yields a complex matrix. Taking the modulus of this matrix gives the S-modulus matrix. The S-transform is used to reveal the time-domain and frequency-domain characteristics of the signal. The S-transform uses a window function to localize the signal in the time domain, and then performs a Fourier transform to obtain its representation in the time-frequency domain. For the input signal, the S-transform is defined as follows: in, Represents the Gaussian window function. As a translation factor, it controls the position of the Gaussian window on the time axis. For frequency, the window width factor is ; The S-transform yields a two-dimensional complex matrix. The S-modulus matrix, obtained by modulo the two-dimensional complex matrix, is then used to select the load characteristic harmonics as the R-channel. The formula for the S-modulus matrix is: in, Let S represent the S-modulus matrix obtained by taking the modulus of a two-dimensional complex matrix.

3. The non-invasive load recognition method based on multi-dimensional feature fusion image according to claim 1, characterized in that, In step one, the power factor-weighted voltage-reactive power trajectory is differentiated by acquiring current data. This involves capturing one cycle of current and voltage data from the voltage zero-crossing point under steady-state conditions. Then, based on Fryze power theory, the load current is decomposed into active and reactive current. The formula for calculating the reactive current is: in, Reactive component of load current Original load current, The active component of the load current, active current The calculation formula is: in, The active power representing the load. This indicates the number of data points within a period. and Each of the following is the first in the period Each current value and voltage value, This is the effective value of the voltage; After obtaining the voltage and reactive current data for the next cycle in steady state, they are normalized, and the calculation formula is as follows: in, and These represent the minimum and maximum values ​​of reactive current during the period, respectively. and These represent the minimum and maximum voltage values ​​during the period, respectively. Create a VI using normalized data. f The trajectory is covered using an M×M zero matrix. Then, all sampling points are traversed. If a sampling point is within the matrix range, the matrix cell value is set to 1. Finally, a representative VI is constructed. f The grayscale matrix of the trajectory; Power factor is an important characteristic of a load, representing the load's efficiency in utilizing the active power of the power source. The power factor is calculated as follows: in, This is the effective value of the current. Indicates apparent power. Indicates the load power factor, VI f The gray value matrix of the trajectory multiplied by Obtain the power factor weighted VI f Trajectory matrix.

4. The non-invasive load recognition method based on multi-dimensional feature fusion image according to claim 1, characterized in that, The feature fusion color image in step two refers to merging different feature information to form a color image with comprehensive features. The feature fusion method is weighted summation, local region merging, and multi-scale fusion. The feature fusion color image provides a more comprehensive and accurate image description, providing more information for image analysis and application.

5. The non-invasive load recognition method based on multi-dimensional feature fusion image according to claim 1, characterized in that, The ImageNet dataset mentioned in step three is a large-scale image database used for computer vision research and algorithm evaluation. The ImageNet dataset contains millions of images from various fields, including a variety of objects, scenes and concepts.

6. The non-invasive load recognition method based on multi-dimensional feature fusion image according to claim 1, characterized in that, The ResNet-18 model in step three is a deep residual network architecture that employs the idea of ​​residual learning. It addresses the degradation in deep neural network training by introducing residual connections. These residual connections allow the network to skip a layer during information flow, enabling the network to learn features better.

7. The non-invasive load recognition method based on multi-dimensional feature fusion image according to claim 1, characterized in that, The transfer learning in step three utilizes previously learned knowledge and experience to improve learning performance on a new task.

Images