A power load identification method based on data enhancement and multi-type feature extraction

Abstract

This invention discloses a power load identification method based on data augmentation and multi-type feature extraction, comprising: extracting transient time series data from historical one-dimensional load signals, mapping it to a two-dimensional probability distribution matrix using the Markov transfer field algorithm, and converting it into a grayscale image; then using random brightness enhancement to simulate light and shadow fluctuations to augment the image data, while combining wavelet transform for image denoising preprocessing; extracting texture and edge features of the MTF image using Gabor transform and LBP operator respectively, inputting them in parallel into a dual-branch convolutional neural network for deep feature extraction, and then completing nonlinear mapping and dimensionality reduction optimization through a multi-layer fully connected network; finally obtaining the load identification model after classification training. This invention has the advantages of strong noise resistance, high generalization ability, and excellent classification accuracy, and can be widely used in the fields of non-intrusive load monitoring and demand-side refined energy management in the context of new power systems.

Description

Technical Field

[0001] This invention relates to the application of deep learning in the field of load identification, specifically to a power load identification method based on multi-source feature fusion and dual-branch convolutional neural network. Background Technology

[0002] Against the backdrop of building a new power system, load identification technology has become a key support for achieving efficient utilization of power resources and supply-demand coordination, in order to promote refined energy management on the demand side and deeply explore the flexible dispatch potential of end users. Non-intrusive load monitoring (NILM), a core technology first proposed by Hart, eliminates the need for installing measuring equipment at each user end. It relies solely on electrical data from the main incoming line to analyze equipment energy consumption information in a low-cost and non-intrusive manner. This not only effectively fills the gap in refined electricity consumption data on the residential side but also deeply activates ultra-short-term dispatchable resources on the user side, providing core support for the coordinated interaction of "source-grid-load-storage". Furthermore, facing massive amounts of electricity consumption data in real-world scenarios, automated load identification technology can accurately capture electricity anomalies, widely empowering multi-dimensional applications such as demand response, energy management, equipment health monitoring, and fire early warning. Therefore, seeking more efficient and automated load identification methods has irreplaceable practical significance for promoting the intelligent upgrading and green, low-carbon transformation of the power system.

[0003] The current achievements in power load identification based on traditional machine learning and deep learning are as follows: Reference 1 (Yanzhen L, Haixin W, Junyou Y, et al. A non-intrusive load monitoring algorithm based on multiple features and decision fusion[J]. Energy Reports, 2021, 7(S7): 1555-1562. DOI: 10.1016 / J.EGYR. 2021.09.087.) proposes a novel NILM method, which achieves deep feature fusion and classifier model fusion through an improved Dempster–Shafer (D–S) evidence theory. First, power features, current harmonic features, and voltage-current (V–I) trajectory features are extracted from the input signal. Then, K-nearest neighbors (KNN), random forest (RF), and convolutional neural network (CNN) are used to identify load devices, utilizing three independent features. Finally, the probability estimation data of each classifier is sent to the aggregator for aggregation to obtain the final identification result of the improved D-S evidence theory; Reference 2 (Ding Maomao, Duan Junxiang, Liu Bo, et al. Anomaly identification of air conditioning load data based on user feature extraction [J]. Electronic Design Engineering, 2026, 34(03):79-82+88.DOI:10.14022 / j.issn1674-6236.2026.03.017.) comprehensively uses statistical, linear regression and other methods to extract user features of air conditioning load data, and uses these features as input information to train a probabilistic neural network (PNN) model for anomaly identification. Through the collaborative processing of the pattern layer, summation layer and decision layer in the model, the effective identification of anomalies in air conditioning load data is realized; Reference 3 (Liu Y, You W, Yang M. Non-Intrusive Load Identification Based on Multivariate Features and Information Entropy-Weighted Ensemble[J].Energies,2025,18(9):2369-2369.DOI:10.3390 / EN18092369.) proposes a non-intrusive load identification method based on multivariate features and information entropy weighted integration. First, one-dimensional numerical features related to power and current are input into a traditional machine learning model, and two-dimensional image features of binary V-I trajectories are processed by the deep neural network model Swin Transformer. Then, the weights of each classification model are adaptively determined using information entropy, and a weighted voting strategy is used to combine the decisions of multiple models to obtain the final identification result.

[0004] While the aforementioned methods based on traditional machine learning and deep learning have made some progress in the field of power load identification, they still reveal a series of shortcomings that urgently need to be addressed in practical applications. First, traditional machine learning relies on human experience in feature engineering and lacks adaptive capabilities. Second, deep learning models based on one-dimensional time series are prone to destroying local waveform details during convolutional sliding processing. Finally, traditional VI trajectory image-based methods have stringent requirements for high-frequency sampling hardware, making deployment difficult and costly. Therefore, there is still a need to seek more efficient and automated load identification methods.

[0005] To address the shortcomings of existing research in feature engineering, this invention proposes a method, system, device, and storage medium for power load identification based on multi-source feature fusion and a dual-branch convolutional neural network. First, an MTF (Mean Transformer Forecasting) graph is constructed based on the original load data. To address the problem of image data imbalance, this invention employs a random brightness augmentation method. While preserving the original image data, the brightness of the image is randomly altered, forcing the sample size of each data category to be expanded to a fixed value. This method simulates light and shadow fluctuations, allowing the CNN to ignore the absolute brightness of the image and learn truly discriminative features. Second, the algorithm of this invention, through wavelet transform and Gabor transform, can extract local features of the image at different scales and directions, effectively addressing the preprocessing problems of complex datasets. These transforms can reduce the impact of noise while preserving important information, laying a solid foundation for subsequent feature extraction. Finally, combined with a CNN convolutional network, leveraging its end-to-end learning capabilities, it can not only automatically learn and extract high-level image features but also optimize the model's classification performance through training on large amounts of data. This combination makes the entire recognition process more automated and efficient, reduces the need for manual intervention, and enables the processing of complex image datasets, improving the model's generalization ability and recognition accuracy. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide a power load identification method based on data augmentation and multi-type feature extraction, which improves the accuracy, robustness and generalization ability of load identification.

[0007] The specific plan is as follows:

[0008] A power load identification method based on data augmentation and multi-type feature extraction, the main steps of which include:

[0009] Step 1: Construction of load feature image based on Markov transfer field (MTF)

[0010] Step 1-1: Set the window length, extract the original one-dimensional load current time series data, and select a typical transient time window that includes the process from equipment startup to steady state;

[0011] Step 1-2: Calculate the numerical distribution of the selected sequence and divide it into multiple quantile intervals according to the size, with each interval representing a Markov state;

[0012] Steps 1-3: After mapping each sampling point in the sequence to the corresponding state interval, the transition frequency between adjacent states is counted along the time axis and the state transition probability is calculated. The two-dimensional expansion is performed strictly according to the time order of the original sequence to construct the Markov transition field probability matrix.

[0013] Steps 1-4: The Markov transition field probability matrix is ​​reduced in dimension by using the piecewise aggregation approximation (PAA) technique to obtain the arithmetic mean of the local block elements and output the compressed two-dimensional probability matrix. Then, the probability matrix is ​​scaled proportionally to the target grayscale range using a linear mapping and converted into a single-channel grayscale image of a specified resolution.

[0014] Step 2: Load image augmentation and wavelet denoising preprocessing based on random brightness enhancement

[0015] Step 2-1: Convert the original unsigned integer data format of the initial single-channel grayscale image into a floating-point data format;

[0016] Step 2-2: Generate a random brightness increment within a preset range, and superimpose the brightness increment onto each pixel value of the image channel to simulate the light and shadow fluctuations in the actual environment;

[0017] Steps 2-3: Traverse the image matrix after superimposing the brightness increment, force the pixel values ​​less than 0 to be set to 0, the pixel values ​​greater than 255 to be set to 255, and keep the values ​​between 0 and 255 unchanged. Then restore the image to 8-bit unsigned integer data. Repeat the above process until the number of samples for each load category is expanded to the preset uniform scale to construct the enhanced load image dataset.

[0018] Steps 2-4: Use two-dimensional wavelet transform to perform multi-level decomposition on the enhanced load image, decomposing the image into approximation coefficients containing structural and low-frequency components, and detail coefficients containing high-frequency details and noise components of different scales and directions.

[0019] Steps 2-5: Set the noise reduction threshold and use a soft thresholding function to perform threshold evaluation and suppression processing on the high-frequency detail coefficients in each direction after decomposition, so as to achieve a balance between noise filtering and local edge signal preservation.

[0020] Steps 2-6: Using the unprocessed low-frequency approximation coefficients and the high-frequency detail coefficients after soft thresholding suppression, the spatial domain image is reconstructed through inverse wavelet transform to generate a denoised load feature image with smooth background high-frequency fluctuations and complete preservation of core topological features.

[0021] Step 3: Image feature extraction of the load image

[0022] Step 3-1: Extract texture features: Convolve the denoised MTF image using a Gabor filter to generate the corresponding texture feature image;

[0023] Step 3-2: Extract edge features: Set the neighborhood window size of Local Binary Pattern (LBP), select the gray value of the center pixel of the window as the threshold, compare it with the gray values ​​of the surrounding neighboring pixels one by one to generate a binary sequence, convert it into LBP feature value, and generate the corresponding edge feature image.

[0024] Step 4: Depth feature extraction and classification training of the load image

[0025] Step 4-1: Deep feature extraction: Input the texture feature image and the edge feature image in parallel into the two branches of the dual-branch convolutional neural network. The two branches are sequentially processed through multiple convolutional blocks for deep feature extraction. After compression by the adaptive average pooling layer, the vectors are flattened into one-dimensional vectors and then nonlinearly combined through fully connected layers to independently output feature vectors representing high-level texture information and edge information.

[0026] Step 4-2: Intermediate layer feature fusion: In the intermediate layer stage of the network, the feature vectors output by the first branch and the second branch are deeply concatenated along the feature dimension to construct a new fused feature vector that simultaneously contains high-level texture features and edge features;

[0027] Step 4-3: Fully connected dimensionality reduction and classification: The newly fused feature vector is input into the subsequent fully connected layer for nonlinear mapping and dimensionality reduction, and finally the probability distribution of various labels is output to obtain the final load identification result.

[0028] Compared with the prior art, the significant technical effects of the present invention are as follows:

[0029] (1) Significantly enhanced noise robustness and model generalization ability: This invention abandons traditional expansion methods such as rotation and flipping, and starts from the physical mapping properties of MTF images, introducing random brightness mapping enhancement technology and combining it with wavelet denoising preprocessing. Without destroying the core topological structure of the loaded MTF image, this technology effectively expands the minority class samples by simulating light and shadow fluctuations, which not only alleviates the overfitting problem caused by unbalanced data distribution, but also uses wavelet transform to smooth high-frequency background noise, providing the model with clean and diverse inputs, and significantly enhancing the anti-interference ability and generalization performance in unknown environments.

[0030] (2) Enhanced multi-dimensional feature extraction capability: This invention introduces Gabor transform and LBP operator to construct a multi-source feature extraction mechanism, breaking the limitation that single features are prone to cognitive blind spots or loss of key visual information. By using Gabor transform to capture global macroscopic texture information of multiple scales and directions, and using LBP operator to finely depict the local microscopic spatial structure and edge details of the image, the two complement each other, transforming a single probability distribution map into a multi-dimensional behavioral fingerprint description, improving the recognition and feature expression capability of devices with similar electrical characteristics.

[0031] (3) Feature fusion optimization and classification accuracy improvement: This invention adopts a mid-layer feature splicing and fusion architecture of a dual-branch convolutional neural network. This architecture design effectively avoids semantic conflicts and spatial noise caused by direct mixing of low-level pixels in the front-end fusion, and also overcomes the problem of missing deep feature interactions caused by simple probability addition in the back-end fusion. This invention extracts the depth of texture and edge features through dual-branch multi-layer convolution and then splices them, enabling the model to fully learn the complex nonlinear collaborative relationship between edges and textures, and showing excellent performance in actual load recognition tasks. Attached Figure Description

[0032] Figure 1 is a general block diagram provided in an embodiment of the present invention;

[0033] Figure 2 shows an example of the expanded dataset in an embodiment of the present invention;

[0034] Figure 3 is a comparison of wavelet transform denoising before and after in an embodiment of the present invention; the left image is the load data image without denoising, and the right image is the preprocessed load data image after wavelet transform denoising.

[0035] Figure 4 shows a comparison before and after Gabor transformation in an embodiment of the present invention; the left image is the image to be processed before image features are extracted without Gabor transformation, and the right image is the feature image after texture features are extracted using Gabor transformation;

[0036] Figure 5 is a comparison of LBP transformation before and after in an embodiment of the present invention; the left image is the load image to be processed before feature extraction by the LBP operator, and the right image is the feature image after edge feature extraction by the LBP operator;

[0037] Figure 6 is a schematic diagram of the dual-branch CNN feature splicing network architecture used in the embodiment of the present invention;

[0038] Figure 7 is a flowchart of the feature splicing algorithm framework in an embodiment of the present invention;

[0039] Figure 8 shows the loss function curves of the model training and validation process in this embodiment of the invention; the left figure shows the initial stage of model training, and the right figure shows the later stage of model training.

[0040] Figure 9 is a confusion matrix diagram of the test set in an embodiment of the present invention. Detailed Implementation

[0041] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0042] Example 1

[0043] Please refer to Example 1 Figure 1 , Figure 1 This is an algorithm block diagram of an electricity load identification method based on data augmentation and multi-type feature extraction disclosed in an embodiment of the present invention. Wherein, Figure 1 The described load identification method is applied in power systems, such as for refined load identification and scheduling under complex operating conditions and small sample scenarios. This invention does not limit its application. Figure 1 As shown, this method can be divided into three key steps: load image enhancement and preprocessing, load image feature extraction, and load image depth feature extraction and classification. Specifically, it can include the following operations:

[0044] S1. Dataset Preparation and Processing: The dataset used in this solution comes from the WHITED global household and industry instantaneous energy dataset. The WHITED dataset records the first 5 seconds of startup for 47 different types of devices at a frequency of 44.1 kHz (224,910 rows of data per record). The first 0.5 seconds (approximately the first 22,000 rows) are sufficient for almost all devices to complete the "off-start-stable" process. After reaching a stable state, the differences in voltage and current among various devices are very small. Therefore, this solution uses the measurement data from the first 0.5 seconds for analysis. In the WHITED dataset, the number of record files for each device varies. Therefore, this solution selects 8 device categories with a relatively large number of records and constructs MTF charts using their current data.

[0045] Load Image Enhancement and Preprocessing: Image enhancement of the load image data is performed using brightness augmentation techniques. Pixel values ​​in the load image are adjusted by adding a brightness increment to the brightness value of each pixel. For each image, its brightness is randomly adjusted to a number within the range [-40, 40] until the target value is reached. The original load image data is then converted to floating-point data to avoid overflow or precision loss during brightness adjustment. Pixel values ​​in each image channel are adjusted. During adjustment, a brightness increment is added to the brightness value of each pixel.

[0046] The specific implementation process of step S1 is as follows:

[0047] S1.1 Dataset Preparation and Processing: First, within a given one-dimensional load voltage time series window of length N... In this invention, the numerical distribution of the sequence is calculated and divided into Q quantile intervals according to their magnitude. In this invention, Q is the default value of 8, and each interval represents a Markov state. After mapping each sampling point in the sequence to its corresponding state interval, the frequency of transitions between adjacent states is counted along the time axis, and the transition from state to state is calculated. Transition to state The probability of generating a size of , thus generating a . The Markov transition matrix W, whose elements Can be defined as

[0048]

[0049] To overcome the drawback of traditional transition matrices losing the temporal dependencies of the original sequence, this scheme expands the spatial transition probabilities strictly according to the temporal order of the original sequence X in a two-dimensional manner, constructing a matrix of size... The Markov transition field matrix M. The elements of this matrix...

[0050]

[0051] in This represents the transition probability between the state corresponding to the i-th point and the state corresponding to the j-th point in the sequence, thus strictly preserving the dynamic evolution characteristics of the load signal. Since the original matrix is ​​too large, this scheme further employs the piecewise aggregation approximation (PAA) technique to reduce its dimensionality, uniformly dividing the original matrix into... Divide the data into local blocks and calculate the arithmetic mean of the elements within each block, then output the compressed two-dimensional probability matrix. Finally, a linear transformation is used to... The probability distribution in the model is scaled proportionally to the range of 0-255, thus formally transforming it into... An 8-bit single-channel grayscale image of resolution is used as input into a subsequent deep learning model for feature extraction.

[0052] S1.2 Load Image Enhancement and Preprocessing: Assuming the original pixel value of the load image is `pixel_value`, the adjusted image's pixel value should be `pixel_value + brightness_factor`, thus changing the overall image brightness. Since the original image is read as a `uint8` data type, directly adding values ​​in `uint8` format—for example, adding `brightness_factor = 40` to a pixel value of 250—won't result in 290; instead, it would overflow and wrap back to 34. This would cause areas that should have been white to suddenly become black noise. Therefore, this invention first converts it to a floating-point number (`float32`) before adding or subtracting, ensuring that 250 + 40 safely becomes 290.0. Subsequently, the `np.clip` function iterates through each pixel value in the entire image matrix. If the calculated pixel value is less than 0, it is forcibly set to 0; if the calculated pixel value is greater than 255, it is forcibly set to 255; if the value is exactly between 0 and 255, it remains unchanged. Finally, restoring the brightness-adjusted image to an 8-bit unsigned integer data type ensures that the adjusted pixel values ​​are limited to between 0 and 255. Figure 2 This describes the image enhancement results obtained by brightness augmentation of a Kettle-type load image. We observe that the same original load image, after undergoing brightness shifts of varying random magnitudes, generates new samples with different global brightness characteristics. The filename suffix _b-39 indicates that the overall pixel brightness has decreased, and _b+28 indicates that the overall pixel brightness has increased. Although the overall brightness distribution of the augmented image exhibits significant random fluctuations, the essential topological structure of the kettle (Kettle) under the MTF feature image is perfectly preserved. Specifically, the bright matrix color block in the upper left corner representing the transient state of device startup, and the uniform, fine grid texture representing the global sinusoidal steady-state operation, remain clearly discernible in all augmented samples without being disrupted.

[0053] In order to preserve the true texture edges of the image and remove high-frequency noise, wavelet denoising processing is required before traditional image feature extraction.

[0054] Assuming the original image is I(x, y), after a two-dimensional wavelet transform, the image can be decomposed into subbands of different scales and orientations. Let the decomposition scale be j, then the wavelet domain of the noisy image can be represented as:

[0055]

[0056] in is a scaling function, representing low-frequency components; Let be a wavelet function, representing the scale. ,direction High-frequency details; These are approximate coefficients; The detail factor is D; D represents the horizontal, vertical, and diagonal directions. This represents additive noise, which is mainly distributed in the high-frequency detail coefficients.

[0057] Noise components can be suppressed by thresholding the detail coefficients. Let the threshold function be... The noise reduction process is as follows:

[0058]

[0059] By adjusting the threshold The threshold function can control which coefficients are retained and which are suppressed, thereby achieving a balance between noise removal and signal preservation. In this invention, the threshold... The threshold value is 5, the threshold function is a soft threshold, the wavelet basis function is a Haar wavelet, and the decomposition level is 1.

[0060] During the reconstruction process, the inverse wavelet transform can be used to reconstruct the denoised image O(x,y):

[0061]

[0062] in These are the detail coefficients after thresholding. This indicates the deviations that may be introduced during the reconstruction process.

[0063] like Figure 3 As shown, the left side is the load data image without denoising, and the right side is the preprocessed load image after denoising using wavelet transform. The comparison reveals a high degree of consistency in the macroscopic visual appearance between the images before and after denoising. This is because the excellent localization analysis capability of wavelet transform perfectly preserves the highly regular core topological structure of the Kettle load due to its purely resistive electrical characteristics. Although the macroscopic visual differences are not significant, at the microscopic level of the feature space, wavelet denoising effectively smooths high-frequency numerical fluctuations in the background region caused by equipment noise or computational rounding. This processing avoids over-sharpening or distortion of the image, preserving key electrical evolution details while providing cleaner numerical input features for subsequent convolutional neural networks (CNNs), effectively reducing the risk of model fitting invalid noise.

[0064] S2. Image feature extraction of the load image: The texture features and edge features of the load image are extracted using Gabor transform and LBP operator, respectively.

[0065] The specific implementation process of step S2 is as follows:

[0066] S2.1, Load image feature extraction based on Gabor transform: Perform Gabor transform on the load image:

[0067]

[0068] in,

[0069]

[0070] This is a Gaussian function, called the window function, where a>0, b>0, and parameter b is used to move the window horizontally. In the above formula, It can also be seen as and Convolution, i.e.

[0071]

[0072] In image processing, two-dimensional Gabor filters are commonly used to extract texture features. For images The real part of a two-dimensional Gabor filter is:

[0073]

[0074] in , , Indicates wavelength. It is the direction angle. Let the standard deviation be the Gaussian envelope. The aspect ratio is the spatial dimension (usually taken as 1).

[0075] In this scheme, the direction parameter The angles are 0°, 45°, 90°, and 135° to determine the filter's response to different texture directions; the frequency is selected based on the texture period of the image. A value of 0.2 or 0.3 is used to capture details at different frequencies in the image; wavelength It is inversely proportional to frequency and matches the size of details in the image. If we choose 0.2, then =1 / f=5, wavelength directly affects the spatial resolution of the filter; standard deviation The width of the Gaussian envelope function is controlled, and 1.0 is usually chosen as the standard deviation, which directly affects the filter's ability to capture texture details.

[0076] like Figure 4As shown, the left image depicts the load image before Gabor transform extraction, while the right image depicts the feature image after Gabor transform extraction of texture features. The essence of Gabor transform is to extract information from different directions and frequencies in an image and convert it into a spatial frequency domain representation. By extracting features from the Gabor-transformed image, a texture feature image can be obtained.

[0077] S2.1 Feature Extraction Based on LBP Method: Define a 3×3 neighborhood window in the image. Take the gray value of the center pixel within the window as a threshold. Compare the gray values ​​of the 8 neighboring pixels with this threshold. If the neighboring pixel value is greater than the center pixel value, the comparison result is 1; otherwise, it is 0. After comparing the 8 pixels in the neighborhood, an 8-bit binary number is obtained. Arranging these bits sequentially yields the LBP value of the center pixel. The LBP feature value reflects the edge information of the center pixel and its neighborhood.

[0078]

[0079] Where P is the number of neighboring pixels, ( , () represents the center element in a 3x3 neighborhood, and its pixel value is , This represents the value of other pixels in the neighborhood. s(x) is the sign function, defined as follows:

[0080]

[0081] like Figure 5 As shown, the left image represents the load image before feature extraction using the LBP operator, while the right image represents the feature image after edge feature extraction using the LBP operator. It can be seen that after LBP processing, each pixel in the original image is replaced with a binary pattern of its local neighborhood pixel values. This binary pattern reflects the features of the edges surrounding that pixel. In different regions and edge parts of the image, LBP can provide detailed and rich feature representations, facilitating subsequent model training.

[0082] S3. Deep Feature Extraction and Classification of Load Images: The texture feature image processed by Gabor transform and the edge feature image processed by LBP transform are input into a dual-branch CNN network for feature extraction. Each branch extracts its own features, and the feature vectors from the two branches are concatenated to obtain a new feature vector that simultaneously contains the texture and edge features of the original image. This new feature vector is then input into the subsequent classification network. After further nonlinear mapping and dimensionality reduction, the probability distribution of each load label is finally output for model training and testing.

[0083] The specific implementation process of step S3 is as follows:

[0084] S3.1 Dual-branch CNN network feature extraction: The model adopts a dual-branch parallel architecture, receiving texture feature images and edge feature images respectively. Each branch contains three convolutional blocks, one flattening layer, and one fully connected layer. The first two convolutional blocks each consist of a convolutional layer, a batch normalization layer, a ReLU activation layer, a max pooling layer, and a Dropout layer, respectively. The third convolutional block consists of a convolutional layer, a batch normalization layer, a ReLU activation layer, and an adaptive average pooling layer, as shown below. Figure 6 As shown.

[0085] The first convolutional blocks of the two branches are respectively fed into preprocessed Gabor texture feature images and LBP edge feature images, both of which are 64×64 single-channel images. Each first convolutional block contains 32 3×3 convolutional kernels, padded with 1 pixel, and then undergoes batch normalization and ReLU activation to introduce non-linear transformations. Subsequently, a 2×2 pooling window with a stride of 2 max-pooling layer is used for downsampling, followed by Dropout (dropout rate 0.25). After this process, 32 32×32 feature maps corresponding to the inputs are output. Their purpose is to extract primary features from the two types of images and initially reduce the spatial dimensionality of the feature maps.

[0086] In both branches, the feature maps output from the first convolutional block are input into the second convolutional block. Both branches use 64 3×3 convolutional kernels, "same" padding, batch normalization, and ReLU activation. Then, they undergo downsampling via a 2×2 max-pooling layer and Dropout processing, outputting more abstract features for the two image types, each consisting of 64 16×16 feature maps. The purpose of this layer is to further extract deeper features based on the initial features, fuse channel information, and continuously reduce computational cost.

[0087] After the first two convolutional blocks, the two branches input the resulting feature maps into the third convolutional block. Both branches use 128 3×3 convolutional kernels, "same" padding, batch normalization, and ReLU activation. Unlike the first two layers, its pooling layer uses an adaptive average pooling layer (AdaptiveAvgPool2d), which forcibly compresses the spatial size of the feature maps, outputting 128 3×3 feature maps respectively. The purpose of this layer is to extract higher-order abstract features and unify the size of the feature maps, laying the foundation for subsequent operations.

[0088] The dimensionality-reduced feature maps are input into the flattening layers of the two branches, respectively, and output one-dimensional feature vectors of length 1152 (3×3×128) corresponding to the edge feature image and texture feature image. The function of this layer is to convert the spatial features output by the pooling layer into a one-dimensional vector format, which is convenient for subsequent fully connected layers to process.

[0089] The flattened one-dimensional vectors are then fed into their respective first fully connected layers. Each of the two branches' fully connected layers contains 256 neurons, which are then processed sequentially through batch normalization, ReLU activation, and Dropout (with a dropout rate of 0.5), outputting 256-dimensional feature vectors corresponding to edge features and texture features, respectively. This process allows the model to learn the complex relationships between features, achieving an initial transformation from raw features to class-discriminating features.

[0090] S3.2, Intermediate layer feature splicing: such as Figure 7 As shown, the two branches yielded 256-dimensional feature vectors corresponding to the texture and edge feature images, respectively. These vectors were then concatenated along the feature dimensions using the `concatenate` function to form a new 512-dimensional feature vector. This concatenated vector simultaneously and completely contains both the high-level texture and edge features of the original image, completing the intermediate-level feature fusion.

[0091] S3.3 Classification and Discrimination: The 512-dimensional feature vector obtained after fusion is input into the final classification network. First, it passes through a fully connected layer with 256 neurons, followed by batch normalization, ReLU activation, and Dropout operations to perform non-linear mapping and dimensionality reduction on the fused features, outputting a 256-dimensional feature vector. Finally, this vector is input into the last fully connected layer with 8 neurons, corresponding to the 8 load classification categories. Combined with the Softmax mechanism within the cross-entropy loss function, the network directly outputs an 8-dimensional class logistic value vector, thus obtaining the final classification probability distribution and completing the multi-classification task.

[0092] S3.4 Model Training and Environment Configuration: This embodiment was trained using the PyTorch 1.12.0 framework on an NVIDIA GeForce MX 450 GPU and an 11th Gen Intel(R) Core(TM) i5-11320H @ 3.20GHz hardware environment. Cross-entropy loss was used as the loss function, Adam was the optimizer, the learning rate was set to 0.001, the batch size was 32, and the epochs were 60. An early stopping mechanism was introduced (termination occurred if the validation set showed no decrease after 10 consecutive epochs), and a fixed global random seed (Seed=42) was used to ensure reproducibility.

[0093] Model evaluation and comparative analysis: After model training, the training results are analyzed by plotting loss function curves and confusion matrices. For example... Figure 8 As shown, the model's loss value decreases rapidly in the early stages of training, and the loss curve and accuracy curve gradually converge as the number of training epochs increases. Figure 9 As shown, when the confusion matrix was plotted on the test set, except for PowerSupply being occasionally misclassified as LEDLight, the vast majority of other categories achieved a recall rate of 0.95 or even 1.00, which indicates that the method can effectively identify different loads.

[0094] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A power load identification method based on data enhancement and multi-type feature extraction, characterized by the steps of include: Step 1: Constructing a load feature image based on the Markov transfer field MTF; Step 2: Load image enhancement and wavelet denoising preprocessing based on random brightness enhancement; Step 3: Image feature extraction of the load image; Step 4: Depth feature extraction and classification training of the load image.

2. The power load identification method based on data augmentation and multi-type feature extraction according to claim 1, characterized in that, Step 1 includes the following steps: Step 1-1: Set the window length, extract the original one-dimensional load current time series data, and select a typical transient time window that includes the process from equipment startup to steady state; Step 1-2: Calculate the numerical distribution of the selected sequence and divide it into multiple quantile intervals according to the size, with each interval representing a Markov state; Steps 1-3: After mapping each sampling point in the sequence to the corresponding state interval, the transition frequency between adjacent states is counted along the time axis and the state transition probability is calculated. The two-dimensional expansion is performed strictly according to the time order of the original sequence to construct the Markov transition field probability matrix. Steps 1-4: The Markov transition field probability matrix is ​​reduced and compressed using the segmented aggregation approximation PAA technique. The arithmetic mean of the local block elements is obtained, and the compressed two-dimensional probability matrix is ​​output. Then, the probability matrix is ​​scaled proportionally to the target grayscale range using a linear mapping, and converted into a single-channel grayscale image of a specified resolution. 3.The power load identification method based on data augmentation and multi-type feature extraction of claim 2, characterized in that, Step 2, load image enhancement, includes the following steps: Step 2-1: Convert the original unsigned integer data format of the initial single-channel grayscale image to a floating-point data format; Step 2-2: Generate a random brightness increment within a preset range, and superimpose the brightness increment onto each pixel value of the image channel to simulate the light and shadow fluctuations in the actual environment; Steps 2-3: Traverse the image matrix after superimposing the brightness increment, force the pixel values ​​less than 0 to be set to 0, the pixel values ​​greater than 255 to be set to 255, and keep the values ​​between 0 and 255 unchanged. Then restore the image to 8-bit unsigned integer data. Repeat the above process until the number of samples for each load category is expanded to a preset uniform size, thus constructing an enhanced load image dataset.

4. The power load identification method based on data augmentation and multi-type feature extraction according to claim 3, characterized in that, Step 2, wavelet denoising, includes the following steps: Steps 2-4: Use two-dimensional wavelet transform to perform multi-level decomposition on the enhanced load image, decomposing the image into approximation coefficients containing structural and low-frequency components, and detail coefficients containing high-frequency details and noise components of different scales and directions. Steps 2-5: Set the noise reduction threshold and use a soft thresholding function to perform threshold evaluation and suppression processing on the high-frequency detail coefficients in each direction after decomposition, so as to achieve a balance between noise filtering and local edge signal preservation. Steps 2-6: Using the unprocessed low-frequency approximation coefficients and the high-frequency detail coefficients after soft thresholding suppression, the spatial domain image is reconstructed through inverse wavelet transform to generate a denoised load feature image with smooth background high-frequency fluctuations and complete preservation of core topological features.

5. The power load identification method based on data augmentation and multi-type feature extraction according to claim 1, characterized in that, Step 3 includes the following steps: Step 3-1: Extract texture features: Convolve the denoised MTF image using a Gabor filter to generate the corresponding texture feature image; Step 3-2: Extract edge features: Set the neighborhood window size of the Local Binary Pattern (LBP), select the gray value of the center pixel of the window as the threshold, compare it with the gray values ​​of the surrounding neighboring pixels one by one to generate a binary sequence, convert it into LBP feature values, and generate the corresponding edge feature image.

6. The power load identification method based on data augmentation and multi-type feature extraction according to claim 5, characterized in that, Step 4 includes the following steps: Step 4-1: Deep feature extraction: Input the texture feature image and the edge feature image in parallel into the two branches of the dual-branch convolutional neural network. The two branches are sequentially processed through multiple convolutional blocks for deep feature extraction. After compression by the adaptive average pooling layer, the vectors are flattened into one-dimensional vectors and then nonlinearly combined through fully connected layers to independently output feature vectors representing high-level texture information and edge information. Step 4-2: Intermediate layer feature fusion: In the intermediate layer stage of the network, the feature vectors output by the first branch and the second branch are deeply concatenated along the feature dimension to construct a new fused feature vector that simultaneously contains high-level texture features and edge features. Step 4-3: Fully connected dimensionality reduction and classification training: The newly fused feature vector is input into the subsequent fully connected layer for nonlinear mapping and dimensionality reduction, and finally the probability distribution of various labels is output to obtain the final load recognition result.

Images