A method for identifying load components of a transformer area based on image coding technology
By employing image coding technology and generative adversarial network models, the problem of insufficient accuracy in identifying load components in low-voltage distribution transformer areas has been solved, enabling accurate identification of load components in transformer areas and supporting the stable operation and load forecasting of the distribution network.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2024-12-09
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies make it difficult to accurately identify the load component power ratio of distributed power sources at the low-voltage distribution substation level, resulting in insufficient stability of distribution network operation and insufficient accuracy of load forecasting. Furthermore, individual metering equipment is costly and involves user privacy issues.
Image encoding technology is used to convert the power curves of load components into grayscale images. A combined model of generative adversarial networks and convolutional neural networks is used to identify the load components of the transformer area through image processing methods. A weighted loss function of adversarial loss, L1 loss and perceptual loss is generated for model training to achieve accurate identification of load components.
It improves the accuracy of load component identification, reduces dependence on external factors, enables semi-supervised training under conditions of lack of measurement data, is applicable to more practical scenarios, and provides accurate information support for load forecasting and demand response.
Smart Images
Figure CN119693708B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical engineering technology, and in particular to a method for identifying load components in transformer substations based on image coding technology. Background Technology
[0002] With the introduction of the "dual carbon" target and the development of smart grids, the penetration rate of distributed generation in distribution networks is continuously increasing. The randomness and volatility of its output pose a series of challenges to the stable operation of the power system. Since most distributed generation is installed after the meter, the measuring devices can only monitor and acquire net load data, resulting in an inability to accurately identify the specific power proportions of various typical load components within the distribution substation. This makes it difficult to provide accurate information support for load forecasting, demand-side response, voltage control, and other measures. At the same time, installing individual metering equipment for all dispersed load components is impractical, not only due to high costs but also because it raises user privacy concerns. Therefore, it is necessary to solve this problem through feasible algorithms to accurately identify the power proportions of different types of load components directly from the net load at the low-voltage distribution substation level, thereby improving the observability of the distribution network.
[0003] Existing load component identification methods can be mainly divided into two categories: one is based on physical models, which establishes power characteristic models using parameters such as meteorological data and capacity information to calculate the specific power of load components. However, physical models may not be consistent across different distribution systems, and accurate model parameters are difficult to obtain in practice, making this type of method difficult to implement. The other is based on data-driven approaches, which analyze the clustering characteristics of load components based on a large amount of measurement data to help identify the power proportion of load components from the net load. However, both of these methods directly process the measured time series, while high-performance models such as convolutional neural networks are better suited for processing image features. The features extracted from the time series are relatively limited, which is not conducive to further improving the identification accuracy of load components. Based on the idea of image encoding, the power time series is converted into an image, and advanced models in the field of image processing can be used to solve the load component identification problem, achieving accurate identification of load components in the distribution area. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention utilizes image coding technology and improves the image processing model to propose a method for identifying the load composition of distribution areas based on image coding technology, providing accurate information support for load forecasting, demand response and other measures in the distribution network.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for identifying the load distribution of transformer substations based on image coding technology, comprising the following steps:
[0006] (1) Based on the cluster characteristics of load components, the characteristic power curves of various components are determined on a daily basis; the characteristic curves of net load and various load components are encoded into two-dimensional grayscale images using the Gram angle and field method.
[0007] (2) Based on the known measurement data of various components, the synthetic minority class oversampling technique is used to generate similar load component power curves; the generated power datasets are combined as the power curves of load components in the same area, and then superimposed to obtain the net load curve and the power ratio of various load components, and synthesize the training dataset of the area identification model.
[0008] (3) Input the grayscale image of the net load and the grayscale image of the load components into the identification model, and extract the image features through the encoder composed of residual blocks and convolutional layers; fuse the features of the net load and various load components into multiple sets of composite features, and use the separation module composed of convolutional layers and deconvolutional layers to identify the images corresponding to various components respectively; and adopt the framework of generative adversarial network, use the discriminator to judge the authenticity of the identified image, and under the adversarial game between the load component identifier and the discriminator, use the weighted sum of adversarial loss, L1 loss and perceptual loss as the loss function, and update the model parameters by backpropagation based on the real images of the training dataset; until the trained load component identification model is obtained.
[0009] (4) For the day to be identified, encode the actual measured data of the net load of the transformer area and various components into images, input them into the trained load component identification model, and obtain the image corresponding to the sum of the power of the load components in the transformer area; then, encode the image into a power curve to complete the load component identification task of the transformer area.
[0010] Further, step (1) specifically involves: as the number of users increases, the load component clusters exhibit strong regularity and are highly correlated with the sum of the load components in the transformer area. Therefore, the sum of the power curves of the observable load components is selected as the characteristic power curves of each type of component, as expressed below:
[0011]
[0012] Among them, P fl The characteristic curve of the load component; Let N be the power curve of the i-th group of observable load components. l The total number of curves included in the considerable load component;
[0013] Normalize the net load curve and the characteristic curves of the load components to the range [0,1]:
[0014]
[0015] in, The data is for the t-th time point on day d after normalization of the power curve; P d,t Let P be the data of the power curve at time t on day d, and let P be the dataset consisting of all power curves.
[0016] Using the normalized power curve as the angle cosine and the timestamp as the radius, transform to the polar coordinate system:
[0017]
[0018] in, Let i be the power value at point i on the normalized daily power curve. They are respectively In polar coordinates, the polar angle and polar radius, t i The timestamp is N, and the total number of sampling points for the daily power curve is N.
[0019] Calculate the cosine of the sum of the polar angles corresponding to each point on the daily power curve to obtain the Gram angle and field:
[0020]
[0021] Where I is a unit row vector; the diagonal elements of the GASF matrix retain the original power information of the load components, and the remaining elements reflect the time correlation between different sampling points;
[0022] Based on the pixel value range, the range of GASF matrix elements is transformed from [0,1] to [0,255] to represent the pixel values of the load component image, and then the grayscale images of net load and load components are drawn.
[0023] Further, step (2) specifically involves: calculating the similarity between each curve and the remaining curves in the historical dataset of the observable load components in the feature space, and finding the K nearest neighbors of each type of component;
[0024] Among them, the characteristics and variation patterns of load components are characterized by daily characteristic indicators of power curves. Therefore, the daily characteristic indicators are selected as average daily power, maximum daily power, minimum daily power, root mean square error, skewness, and kurtosis, in order to measure the similarity between curves.
[0025] Generating similar load component power curves using minority class oversampling techniques:
[0026]
[0027] in, This represents the power data of the q-th load component generated by user i on day d. The power data of the load component measured by user i on day d is as follows. The similar power curves are selected from the K nearest neighbors, rand(0,1) is a random number between [0,1], and Q is the number of power curves generated;
[0028] Based on the maximum value of the load component, a capacity factor is set to determine the power level of the generating load component:
[0029]
[0030] Where, r l The capacity factor represents the load component. Since the capacity differences between user load components are relatively small, the coefficient range is selected as [0.5, 1.5]; P nl The composite load component curve set is obtained by normalizing the maximum value of the generated curve and then multiplying it by a set capacity value.
[0031] The generated datasets of various components are combined to represent the load components within the same transformer area; the sum of the load components' power and net load data within the transformer area are then superimposed to synthesize the simulation dataset used for model training.
[0032]
[0033] in, These are the combined power of the load components and the net load of the centralized substation area in the simulation training, respectively, at the t-th time point on day d. For the power value of the i-th power curve at time t on day d, N l Number of curves generated; Let M be the power value of the sum of power of the j-th type of load components at time t on day d, and M be the number of load component categories.
[0034] Further, step (3) specifically involves the following: the load component identification model includes a load component identification module and a discrimination module. The load component identification module includes an encoder submodule and multiple heads. The encoder is used to extract image features of the input net load and load components. The heads are used to perform an identification task of a certain type of load component to obtain the image identification result of that type of component. The discrimination module is used to judge the authenticity of the image identification result, thereby learning from the identification module and training the model parameters until a well-trained load component identification model is obtained.
[0035] The encoder submodule comprises three branches, each including multiple convolutional layers and residual blocks. It captures multi-scale image features using convolutional kernels of different sizes and network layers. The encoder submodule extracts features from the input image and inputs them into a head. The number of heads is the same as the number of categories of the payload components, used to output images corresponding to each category. Each head includes a feature fusion module and an image reconstruction module. The feature fusion module concatenates the features of the net payload and corresponding payload components along the channel dimension and reduces the dimensionality to obtain fused features. The image reconstruction module includes multiple convolutional layers and deconvolutional layers, reconstructing images corresponding to single-category payload components based on the fused features.
[0036] The discrimination module receives the load component identification image and the actual image as input, calculates the authenticity of the input image through multiple convolutional layers, and iteratively updates the parameters of the load component identification model in an adversarial game between the identification module and the discrimination module.
[0037] Further, in step (3), a weighted sum of adversarial loss, L1 loss, and perceptual loss is used as the loss function. The model parameters are updated by backpropagation based on the real images in the training dataset. Specifically, during model training, a composite loss function is constructed using the weighted sum of adversarial loss, L1 loss, and perceptual loss, and the load component identification model is updated by backpropagation. Among them, adversarial loss is used to measure the overall quality and realism of the load component images; L1 loss and perceptual loss are used to measure the similarity of the content and texture features of the load component images, respectively. The calculation formula is as follows:
[0038]
[0039] L S (S,D)=λ1L GAN (S,D)+λ2L L1 (S)+λ3L VGG (S);
[0040] Where X is the input image of the load components, S is the identification module, S(X) is the identification result, and Y is the true image of the load components; L S L represents the total loss function of the model, and λ1, λ2, and λ3 are the weight coefficients of the three loss functions. Initially, they are set to equal weights and dynamically adjusted during training. If a certain loss function decreases too quickly, its weight coefficient is reduced. GAN To combat loss, it is possible to measure model performance; L L1 For L1 loss, the mean absolute error is calculated pixel-by-pixel to measure the content loss of the image; L VGG To perceive the loss, a pre-trained VGG-19 network is used to extract high-level features of image texture information, thereby measuring the similarity of images in the feature space; among which, W represents the feature output of the i-th convolutional module at the j-th layer of the VGG network.i,j H i,j For the height and width of the feature, ω i The weights corresponding to the i-th feature group are selected as the weight coefficients used in the pix2pixHD network.
[0041] Further, in step (4), for the day to be identified, the characteristic power curves of the net load and the observable load components are encoded into images and input into the trained load component identification model to obtain the image corresponding to the sum of the load components in the transformer area; subsequently, the identification result of the load components in the transformer area is obtained through the image inverse encoding technique; the expression is as follows:
[0042]
[0043] in, This is the identification value of the load component power curve at point i. Let be the value of the GASF matrix corresponding to the load component image at the i-th point on the diagonal.
[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0045] 1. This invention addresses the limitations of processing measurement data from a time series perspective. It identifies the power proportion of load components in a transformer area through image encoding and image separation, eliminating the need for external factors such as weather and environment. It effectively integrates the characteristics of net load and observable load components, making the identification results of load components more accurate.
[0046] 2. The Grammy angle field algorithm is used to encode the image, which can effectively preserve the temporal correlation of the load, remove redundant information, and give full play to the powerful feature extraction capability of the model.
[0047] 3. Based on the idea of data augmentation, new samples are generated using power data of a small number of load components, and then a training dataset is synthesized. This enables semi-supervised training of the model under the condition of lack of measurement data, so that the load component identification model can be applied to more real-world operating scenarios. Attached Figure Description
[0048] Figure 1 This is a schematic diagram illustrating the implementation framework of the transformer area load component identification method of the present invention.
[0049] Figure 2 This is a schematic diagram of the load component image encoding in this invention;
[0050] Figure 3 This is a schematic diagram of the load component identification module in this invention. Detailed Implementation
[0051] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.
[0052] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0054] This invention, by fully utilizing measurement data of net load and various load components, and combining image coding algorithms and an improved image separation model, proposes an image coding-based method for identifying the load components of distribution substations. This method identifies the power proportion of various load components within a distribution substation, thereby accurately determining the load status of the distribution network and supporting the dispatching and operation of the power grid. Figure 1 The method for identifying load components in transformer substations according to the present invention includes the following steps:
[0055] (1) Based on the clustering characteristics of load components, the characteristic power curves of various components are determined on a daily basis. The characteristic curves of net load and various load components are encoded into two-dimensional grayscale images using the Gram angle and field method.
[0056] First, as the number of users increases, the load component clusters exhibit stronger regularity, and their correlation with the sum of the load components in the distribution area also increases significantly. Therefore, the sum of the power curves of the considerable load components is selected as the characteristic power curves of each component:
[0057]
[0058] Among them, P fl The characteristic curve of the load component; Let N be the power curve of the i-th group of observable load components. l The total number of curves included in the considerable load component.
[0059] Secondly, such as Figure 2 As shown, image encoding of the load component characteristic curve includes the following steps:
[0060] (1.1) Normalize the net load curve and the characteristic curves of the load components to the range [0,1]:
[0061]
[0062] in, The data is for the t-th time point on day d after normalization of the power curve; P d,t Let P be the data of the power curve at time t on day d, and let P be the dataset consisting of all power curves.
[0063] (1.2) Using the normalized power curve as the angle cosine and the timestamp as the radius, transform to the polar coordinate system:
[0064]
[0065] in, Let i be the power value at point i on the normalized daily power curve. They are respectively In polar coordinates, the polar angle and polar radius, t i is the timestamp, and N is the total number of sampling points for the daily power curve.
[0066] (1.3) Calculate the cosine of the sum of the polar angles corresponding to each point on the daily power curve to obtain the Gram angle and field:
[0067]
[0068] Where I is a unit row vector. The diagonal elements of the GASF matrix retain the original power information of the load components, while the remaining elements reflect the temporal correlation between different sampling points.
[0069] (1.4) Based on the pixel value range, the range of GASF matrix elements is transformed from [0,1] to [0,255] to represent the pixel values of the load component image, and then the grayscale images of net load and load components are drawn.
[0070] (2) Based on the known measurement data of various components, synthetic minority class oversampling technique is used to generate similar load component power curves. The generated power datasets are combined as the power curves of load components in the same area, and then superimposed to obtain the net load curve and the power ratio of various load components, thus synthesizing the training dataset of the area identification model.
[0071] First, calculate the similarity between each curve and the remaining curves in the historical dataset of observable load components in the feature space to find the K-nearest neighbors of each component. To ensure accuracy, the value of K should be relatively small. The historical dataset can be divided into training and validation sets and selected using cross-validation.
[0072] Among them, the characteristics and variation patterns of load components are characterized by daily characteristic indicators of the power curve. The daily characteristic indicators are selected as the average daily power, maximum daily power, minimum daily power, root mean square error, skewness, and kurtosis, in order to measure the similarity between curves.
[0073] Subsequently, a similar load component power curve was generated using minority class oversampling techniques:
[0074]
[0075] in, This represents the power data of the q-th load component generated by user i on day d. The power data of the load component measured by user i on day d is as follows. The similar power curves are selected from the K nearest neighbors, rand(0,1) is a random number between [0,1], and Q is the number of power curves generated.
[0076] Subsequently, based on the maximum value of the load component, a capacity factor is set to determine the power level of the generating load component:
[0077] P nl =[P nl / max(P nl )]*r l *max(P ol )r l ∈[0.5,1.5];
[0078] Where, r l This represents the capacity factor of the load components. Since the capacity differences between user load components are relatively small, the coefficient range is selected as [0.5, 1.5]. P nl The composite load component curve set is obtained by normalizing the maximum value of the generated curve and then multiplying it by a set capacity value.
[0079] Subsequently, the generated datasets of various components are combined to represent the load components within the same transformer area. The sum of the load components' power and net load data within the transformer area are then overlaid to synthesize the simulation dataset used for model training.
[0080]
[0081]
[0082] in, These represent the combined power of the load components and the net load in the simulation training area at time point t on day d, respectively. For the power value of the i-th power curve at time t on day d, N l To generate the number of curves. Let M be the power value of the sum of power of the j-th type of load components at time t on day d, and M be the number of load component categories.
[0083] (3) Input the grayscale images of the net load and load components into the identification model, and extract image features through an encoder composed of residual blocks and convolutional layers. Then, the features of the net load and various load components are fused into multiple sets of composite features, and the images corresponding to each component are identified by a separation module composed of convolutional and deconvolutional layers. At the same time, a generative adversarial network framework is adopted, and the discriminator judges the authenticity of the identified images. Under the adversarial game between the load component identifier and the discriminator, the weighted sum of adversarial loss, L1 loss and perceptual loss is used as the loss function, and the model parameters are updated by backpropagation based on the real images in the training dataset.
[0084] The load component identification model includes a load component identification module and a discrimination module. For example... Figure 3 This is a schematic diagram of the load component identification module. The identification module includes an encoder and multiple heads. The encoder is used to extract image features of the input net load and load components. The heads are used to perform identification tasks for a certain type of load component to obtain the image identification result for that type of component. The discrimination module is used to judge the authenticity of the image identification result, thereby learning from the identification module through game theory and training the model parameters.
[0085] like Figure 3 As shown, the encoder submodule comprises three branches, each including multiple convolutional layers and residual blocks. It captures multi-scale image features using convolutional kernels of different sizes and network layers. After extracting features from the input image, the encoder submodule feeds them into the head. The number of heads is the same as the number of payload component categories, used to output images corresponding to each component. Each head includes a feature fusion module and an image reconstruction module. The feature fusion module concatenates the features of the net payload and corresponding payload components along the channel dimension and reduces the dimensionality to obtain fused features. The image reconstruction module includes multiple convolutional and deconvolutional layers, reconstructing images corresponding to single-class payload components based on the fused features.
[0086] The discrimination module receives the load component identification image and the actual image as input, calculates the authenticity of the input image through multiple convolutional layers, and iteratively updates the parameters of the load component identification model in an adversarial game between the identification module and the discrimination module.
[0087] During model training, a composite loss function is constructed using a weighted sum of adversarial loss, L1 loss, and perceptual loss, and backpropagation is used to update the load component identification model. The adversarial loss measures the overall quality and realism of the load component images; the L1 loss and perceptual loss measure the similarity of content and texture features in the load component images, respectively. The calculation formulas are as follows:
[0088]
[0089]
[0090] L S (S,D)=λ1L GAN (S,D)+λ2L L1 (S)+λ3L VGG (S);
[0091] Where X is the input load component image, S is the identification module, S(X) is the identification result, and Y is the true load component image. S L represents the total loss function of the model, and λ1, λ2, and λ3 are the weight coefficients of the three loss functions. Initially, they are set to equal weights and dynamically adjusted during training. If a certain loss function decreases too quickly, its weight coefficient is reduced. GAN To combat loss, it is possible to measure model performance; L L1 For L1 loss, the mean absolute error is calculated pixel-by-pixel to measure the content loss of the image; L VGG To address the perceived loss, a pre-trained VGG-19 network is used to extract high-level features such as image texture information, thereby measuring the similarity of images in the feature space, which better aligns with human visual perception of image differences. W represents the feature output of the i-th convolutional module at the j-th layer of the VGG network. i,j H i,j For the height and width of the feature, ω i The weights corresponding to the i-th feature group are selected as the weight coefficients used in the pix2pixHD network.
[0092] The synthetic training dataset is used as real transformer area measurement data. The load component identification model is used to estimate the power proportion of each component, the loss function between the identification results and the real data is calculated, and the model parameters are continuously updated iteratively.
[0093] (4) For the day to be identified, the actual measured data of the net load and various components of the transformer area are encoded into images and input into the trained identification model to obtain the image corresponding to the sum of the power of the load components in the transformer area. Subsequently, the image is inverse encoded into a power curve to complete the task of identifying the load components in the transformer area; the identification results of the load components in the transformer area are obtained through the inverse encoding technique of the image; the expression is as follows:
[0094]
[0095] in, This is the identification value of the load component power curve at point i. Let be the value of the GASF matrix corresponding to the load component image at the i-th point on the diagonal.
[0096] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.
[0097] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.
Claims
1. A method for identifying the load distribution of transformer substations based on image coding technology, characterized in that, Includes the following steps: (1) Based on the cluster characteristics of load components, determine the characteristic power curves of various components on a daily basis; use the Gram angle and field method to encode the characteristic curves of net load and various load components into two-dimensional grayscale images; (2) Based on the known measurement data of various components, the synthetic minority class oversampling technique is used to generate similar load component power curves; the generated power datasets are combined as the power curves of load components in the same area, and then superimposed to obtain the net load curve and the power ratio of various load components, and the training dataset of the area identification model is synthesized. (3) Input the grayscale image of the net load and the grayscale image of the load components into the identification model, and extract the image features through the encoder composed of residual blocks and convolutional layers; fuse the features of the net load and various load components into multiple sets of composite features, and use the separation module composed of convolutional layers and deconvolutional layers to identify the images corresponding to various components respectively; and adopt the framework of generative adversarial network, use the discriminator to judge the authenticity of the identified image, and under the adversarial game between the load component identifier and the discriminator, use the weighted sum of adversarial loss, L1 loss and perceptual loss as the loss function, and backpropagate the real images of the training dataset. The model parameters are updated via backpropagation until a well-trained load component identification model is obtained. Specifically, the weighted sum of adversarial loss, L1 loss, and perceptual loss is used as the loss function, and the model parameters are updated via backpropagation based on real images in the training dataset. During model training, a composite loss function is constructed using the weighted sum of adversarial loss, L1 loss, and perceptual loss, and the load component identification model is updated via backpropagation. The adversarial loss measures the overall quality and realism of the load component images; the L1 loss and perceptual loss measure the similarity of the content and texture features of the load component images, respectively. The calculation formulas are as follows: ； ； ； ； in, The input is a load component image. For the identification module, To identify the results, This is a true image of the load components; The total loss function of the model, These are the weight coefficients for the three types of loss functions, initially set to equal weights, and dynamically adjusted during training. If a certain loss function decreases too quickly, its weight coefficient is reduced. To combat loss, it is possible to measure model performance; For L1 loss, the mean absolute error is calculated pixel by pixel to measure the content loss of the image; To perceive the loss, a pre-trained VGG-19 network is used to extract high-level features of image texture information, thereby measuring the similarity of images in the feature space; among which, For the VGG network The first convolutional module Features of layer output, The height and width of the feature, For the first The weights corresponding to the group features are selected as the weight coefficients used in the pix2pixHD network. (4) For the day to be identified, encode the actual measured data of the net load of the transformer area and various components into images, input them into the trained load component identification model, and obtain the image corresponding to the sum of the power of the load components in the transformer area; then, encode the image into a power curve to complete the load component identification task of the transformer area.
2. The method for identifying load components in transformer substations according to claim 1, characterized in that, The specific steps (1) are as follows: As the number of users increases, the load component clusters exhibit strong regularity and are highly correlated with the sum of the load components in the transformer area. Therefore, the sum of the power curves of the observable load components is selected as the characteristic power curves of each type of component, and the expression is as follows: ； in, The characteristic curve of the load component; For the first The power curves of the group of observable load components, The total number of curves included in the considerable load component; Normalize the net load curve and the characteristic curves of the load components to the range [0,1]: ； in, The power curve after normalization Heavenly Data at each point in time; For the first Heavenly Data of power curves at each time point, This is a dataset consisting of all power curves; Using the normalized power curve as the angle cosine and the timestamp as the radius, transform to the polar coordinate system: ； in, The normalized daily power curve is on the 1st The power value of the point, They are respectively In polar coordinates, polar angle and polar radius, For timestamps, This represents the total number of sampling points for the daily power curve. Calculate the cosine of the sum of the polar angles corresponding to each point on the daily power curve to obtain the Gram angle and field: ； in, The GASF matrix consists of unit row vectors; the diagonal elements of the GASF matrix retain the original power information of the load components, while the remaining elements reflect the temporal correlation between different sampling points. Based on the pixel value range, the range of GASF matrix elements is transformed from [0,1] to [0,255] to represent the pixel values of the load component image, and then the grayscale images of net load and load components are drawn.
3. The method for identifying load components in transformer substations according to claim 1, characterized in that, The specific steps (2) are as follows: calculate the similarity between each curve and the other curves in the historical dataset of the observable load components in the feature space, and find the K nearest neighbors of each type of component; Among them, the characteristics and variation patterns of load components are characterized by daily characteristic indicators of power curves. Therefore, the daily characteristic indicators are selected as average daily power, maximum daily power, minimum daily power, root mean square error, skewness, and kurtosis, in order to measure the similarity between curves. Generating similar load component power curves using minority class oversampling techniques: ； in, Indicates a considerable number of users The generated first Group load components in the first Daily power data For a considerable number of users The measured loading components in the first Daily power data The similar power curves selected from the K nearest neighbors are... A random number between [0,1] The number of power curves generated; Based on the maximum value of the load component, a capacity factor is set to determine the power level of the generating load component: ； in, The capacity coefficient represents the load component. Since the capacity differences between user load components are relatively small, the coefficient range is selected as [0.5, 1.5]. The composite load component curve set is obtained by normalizing the maximum value of the generated curve and then multiplying it by a set capacity value. The generated datasets of various components are combined to represent the load components within the same transformer area; the sum of the load components' power and net load data within the transformer area are then superimposed to synthesize the simulation dataset used for model training. ； ； in, The simulated training session focused on the load components, power composition, and net load in the central substation area at the [number]th [year]. Heavenly Power values at each time point; For the generated first The group power curve in the first Heavenly Power values at each time point Number of curves generated; For the first The sum of the power of the load components in the first class Heavenly Power values at each time point This represents the number of load component categories.
4. The method for identifying load components in transformer substations according to claim 1, characterized in that, The specific step (3) is as follows: the load component identification model includes a load component identification module and a discrimination module. The load component identification module includes an encoder submodule and multiple heads. The encoder is used to extract image features of the input net load and load components. The heads are used to perform an identification task of a certain type of load component to obtain the image identification result of that type of component. The discrimination module is used to judge the authenticity of the image identification result, and thus learns from the identification module to train the model parameters; until a well-trained load component identification model is obtained. The encoder submodule comprises three branches, each including multiple convolutional layers and residual blocks. It captures multi-scale image features using convolutional kernels of different sizes and network layers. The encoder submodule extracts features from the input image and inputs them into a head. The number of heads is the same as the number of categories of the payload components, used to output images corresponding to each category. Each head includes a feature fusion module and an image reconstruction module. The feature fusion module concatenates the features of the net payload and corresponding payload components along the channel dimension and reduces the dimensionality to obtain fused features. The image reconstruction module includes multiple convolutional layers and deconvolutional layers, reconstructing images corresponding to single-category payload components based on the fused features. The discrimination module receives the load component identification image and the actual image as input, calculates the authenticity of the input image through multiple convolutional layers, and iteratively updates the parameters of the load component identification model in an adversarial game between the identification module and the discrimination module.
5. The method for identifying load components in transformer substations according to claim 1, characterized in that, In step (4), for the day to be identified, the characteristic power curves of the net load and the observable load components are encoded into images and input into the trained load component identification model to obtain the image corresponding to the sum of the load components in the transformer area; subsequently, the identification result of the load components in the transformer area is obtained through the image inverse encoding technique; the expression is as follows: ； in, The load component power curve at the 1st The identification value of a point The GASF matrix corresponding to the load component image is located on the diagonal. The value of the point.