Power data recovery method based on visual perception

Abstract

The invention relates to an electric power data recovery method based on visual perception, which is used for recovering by converting power data into a power curve image and utilizing a visual perception technology. Firstly, data collection and image generation are performed, and a data set is constructed; secondly, feature extraction model training is carried out, then image similarity calculation and discrimination are carried out, finally, image recovery is carried out, and continuity and accuracy of recovery data are ensured. According to the method, the similarity of historical data can be fully mined by adopting a visual perception technology and by means of image feature extraction and comparison, and missing data can be accurately recovered by considering the particularity of wind power generation influenced by meteorological factors. The visual perception model not only can identify images with similar large trends, but also can capture detail features and provide a more accurate recovery scheme, so that the defect that the similarity of large-scale historical data is not considered in an existing method is effectively overcome.

Description

Technical Field

[0001] The present invention relates to the field of data recovery technology, and in particular to a method for recovering missing power data based on visual perception technology, which is applicable to renewable energy fields such as wind power generation, solar power generation, and other scenarios requiring high-precision data recovery. Background Art

[0002] In the fields of wind power generation, solar power generation, etc., data loss is often caused by extreme weather, equipment failure or network problems during the collection of power data. These losses may appear as single-point intermittent loss or continuous small segment loss, which seriously affects the dispatch and prediction accuracy of the power system. Traditional data recovery methods such as numerical interpolation and generative adversarial networks (GAN) have certain limitations when dealing with complex scenarios and nonlinear data loss. The numerical interpolation method only uses the numerical relationship before and after the missing data, and fails to make full use of the trend information in the historical data; the GAN-based recovery method is easily affected by data noise and has poor adaptability to extreme weather or emergencies. In addition, the current recovery requirements also include high-precision recovery of missing data in a complex and changeable environment, while requiring rapid processing of large-scale data and ensuring reliable results. In order to solve the above problems, the present invention proposes a power data recovery method based on visual perception, which converts data into image form and combines deep learning technology to achieve accurate recovery of missing data, significantly improving the recovery efficiency and accuracy. Summary of the invention

[0003] The present invention provides a power data recovery method based on visual perception, which aims to effectively recover the missing parts of wind power data through deep learning technology. The method converts power data into power curve images, uses image feature extraction and visual perception technology, and combines deep learning models to infer the missing power data from the similarity of images.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A power data recovery method based on visual perception comprises the following steps:

[0006] Step 1: Data collection and image generation: based on wind power data collected from multiple wind farms, the time window is defined and converted into a power curve image to construct a data set;

[0007] Step 2: Feature extraction model training, extracting the features of wind power curve through convolutional neural network autoencoder combined with residual network and attention mechanism. Feature extraction only involves the encoding stage of the autoencoder, and the extracted features are saved as independent files for subsequent similarity calculation and data recovery tasks;

[0008] Step 3, image similarity calculation and discrimination: by calculating the similarity between image features, based on the overall trend and fine features of the image, find the image segment that is most similar to the target image;

[0009] Step 4, image restoration: Analyze the missing area in the target image and determine the missing location. Combine the most similar image segments, use image restoration technology to perform smoothing, interpolate missing data, and ensure the continuity and accuracy of the restored data.

[0010] Compared with the prior art, the present invention has the following beneficial effects:

[0011] Existing methods cannot fully consider the trend and similarity of historical data. Traditional methods, such as numerical interpolation and recovery methods based on generative adversarial networks, mainly rely on the numerical relationship between previous and subsequent data to infer missing values, ignoring trend data that may have similar situations in history. This is not effective when dealing with complex weather changes and irregular data missing.

[0012] Existing methods have great limitations. Generative adversarial networks and numerical methods are more based on the characteristics of the data itself. They are not flexible enough and cannot fully utilize the potential patterns of historical data. For example, meteorological conditions such as wind speed in wind power generation usually have similar historical scenes. These patterns can be better captured and utilized through visual perception technology.

[0013] The innovation of this invention lies in restoring data through an image-driven method, using historical images of wind power for similarity matching, and accurately restoring based on the trend and similarity of historical data. This method is different from existing numerical inference methods and relies more on visual perception technology, which can improve the reliability of recovery when processing large-scale data.

[0014] Most of the existing recovery schemes are based on numerical methods or data-driven technologies such as generative adversarial networks. The limitation of these methods is that they do not fully consider the similarity of historical data, and rely more on the numerical laws before and after the missing data for inference. In contrast, the present invention, by adopting visual perception technology, with the help of image feature extraction and comparison, can fully explore the similarity of historical data, taking into account the particularity of wind power generation affected by meteorological factors, and then accurately restore the missing data. The visual perception model can not only identify images with similar general trends, but also capture detailed features to provide a more accurate recovery solution, thereby effectively solving the deficiency of existing methods that fail to consider the similarity of large-scale historical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the process of the present invention.

[0016] Figure 2It is a schematic diagram of wind power generation curve (model training input data).

[0017] Figure 3 It is a schematic diagram of wind power generation curve (including missing values).

[0018] Figure 4 It is the experimental result of similar image search without missing images.

[0019] Figure 5 It is the similar image search and data recovery result of continuous multi-point missing images. DETAILED DESCRIPTION

[0020] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0021] refer to Figure 1 , the complete process of the present invention is as follows:

[0022] Step 1: Data acquisition and image generation.

[0023] This step collects wind power data from multiple wind farms and converts these time series data into power curve images. This step provides input data suitable for deep learning model processing for subsequent data recovery tasks (such as Figure 2 The workflow of this step includes data acquisition, time window division, power curve drawing and image generation.

[0024] The time interval for wind power data collection is once an hour, so the data for each day includes power values ​​at 24 time points. In addition to power data, the system does not involve other environmental factors or meteorological conditions, and focuses on the temporal changes of power data. In order to fully explore the laws of power data and expand the data set, this step not only adopts the traditional time window (i.e., from 00:00 to 00:00 the next day), but also introduces a sliding time window strategy. Through this strategy, additional power images are generated in different time periods. The specific sliding time windows include: from 06:00 to 06:00 the next day, from 12:00 to 12:00 the next day, and from 18:00 to 18:00 the next day. In this way, the model can learn different power generation laws from power data in multiple time periods, thereby enhancing the diversity and richness of the data. The power images generated in these different time periods will be used as independent data samples for subsequent image recognition and missing data recovery tasks.

[0025] Data analysis shows that data loss may occur due to equipment failure or extreme weather conditions. Missing data may appear as single-point intermittent loss or small-segment continuous loss (such as Figure 3 Figure 2.1 shows the missing value data).

[0026] Step 2: feature extraction and model training.

[0027] The feature extraction model training step builds and trains a deep learning architecture that integrates a convolutional neural network autoencoder and a residual network, which is specifically used to extract deep features of wind power curve images. The core of this step is the encoding stage of the autoencoder, which compresses high-dimensional image data into low-dimensional feature representations and saves these features as independent files for subsequent similarity calculations and data recovery tasks.

[0028] During the data preprocessing process, the input wind power curve image is first uniformly adjusted to 224×224 pixels to meet the input requirements of the model. Subsequently, a series of data enhancement techniques are applied, including random horizontal flipping, translation, scaling, and rotation. These operations are implemented through a Python library to increase the diversity of training data and improve the generalization ability of the model. Normalization is based on the mean and standard deviation commonly used in pre-trained models, that is, the mean is (0.485, 0.456, 0.406) and the standard deviation is (0.229, 0.224, 0.225) to accelerate the convergence process of the model. The specific preprocessing and enhancement steps are as follows:

[0029] x aug =Normalize(Transform(x))

[0030] Where x is the original input image, x aug The image is after data augmentation and normalization.

[0031] The encoder part consists of multiple convolutional layers, batch normalization layers, ReLU activation functions, residual blocks, and attention mechanisms. The operation process of each convolutional layer is as follows:

[0032] H (l) =RELU(BatchNorm(W (l) *H l-1 +b (l) ))

[0033] Among them, H (l-1) is the input feature map of the l-1th layer, W (l) is the convolution kernel weight of the lth layer, b (l) is the bias term, ReLU is the activation function, * represents the convolution operation, H (l) is the output feature map of the lth layer.

[0034] Subsequently, the residual block is introduced to enhance the feature extraction capability and alleviate the gradient vanishing problem in deep networks. The operation process of the residual block is as follows:

[0035]

[0036] y=F(x)+W s (l) *x

[0037] Among them, x is the input feature map, W1 (l) and are the weight matrices of the two convolutional layers in the lth residual block, and is the corresponding bias vector, W s (l) The convolution kernel used to adjust the number of channels ensures that the dimensions of the jump connection match. The jump connection directly adds the input feature map to the output after the convolution transformation, achieving an identity mapping and enhancing the efficiency of information transmission.

[0038] After each residual block, this method introduces an attention mechanism to further improve the effect of feature extraction. The attention mechanism calculates the importance weights of different regions in the feature map and dynamically adjusts the response strength of each channel of the feature map, so that the model can pay more attention to the key information in the image.

[0039] In the attention mechanism, the feature map undergoes a global average pooling operation to obtain the global features of each channel:

[0040]

[0041] Furthermore, the attention weights are generated through a fully connected layer and a Sigmoid activation function:

[0042] A=σ(W a ·z+b a )

[0043] Among them, W a is the weight matrix of the fully connected layer, b a is the bias vector, σ is the Sigmoid activation function, and A is the attention weight vector. Next, the attention weight is applied to each channel of the feature map:

[0044]

[0045] in, represents an element-wise multiplication operation, is the output feature map of the residual block, is the feature map adjusted by the attention mechanism. After each residual block and attention mechanism, the maximum pooling layer is applied to further reduce the spatial dimension of the feature map. Through multiple convolutions, batch normalization, ReLU activation, residual blocks, attention mechanism and maximum pooling operations, the encoder gradually reduces the spatial dimension of the input image while increasing the number of feature channels, and finally maps the input image to a high-dimensional feature space. The generated feature representation is:

[0046] z=ε(x;θ ε )

[0047] Where x is the input image, ε represents the encoder function, and θ ε is the encoder parameter, and z is the extracted feature vector. During the feature extraction process, the encoder's output feature vector z is expressed by the following formula:

[0048] f i =ε(x i θ ε )

[0049] Among them, f i is the feature vector of the i-th image. All feature vectors are organized into a matrix:

[0050]

[0051] Among them, M is the total number of images and D is the dimension of the feature vector. The feature vector captures the high-level detail features in the wind power curve image through the encoder's multi-layer convolution, residual network and attention mechanism structure, and the final features are saved in a special type of independent file. Through the above method, the feature extraction model training realizes the conversion of high-dimensional wind power curve images into low-dimensional feature vectors, which provide the basis for subsequent similarity calculation and data recovery tasks.

[0052] Step 3: Image similarity calculation and discrimination.

[0053] Image similarity calculation and discrimination are achieved through deep learning technology, combined with convolutional neural network decoder, residual network and VGG19 perception features, to achieve accurate similarity evaluation of wind power curve images and high-quality recovery of missing data. First, for the target image, the same preprocessing steps as in the training phase are applied, including resizing, data enhancement and normalization, and then it is input into the encoder to generate the feature vector f of the target image. t , its mathematical expression is:

[0054] f t =ε(x t θ ε )

[0055] Among them, x t is the target image, ε represents the encoder function, θ ε is the encoder parameter, f t is the feature vector of the target image. Next, by calculating the feature vector of the target image f t With all image feature vectors f in the database i The cosine similarity between the two images is used to quantify the degree of similarity between the images. The formula for calculating cosine similarity is as follows:

[0056]

[0057] Among them, f t ·f i Represents the vector f t With f i The dot product of ||f t ||2 and ||f t ||2 are vectors f t and f i By calculating all similarity values, we can identify the image x that is closest to the target image in the feature space. best , and its similarity score is:

[0058]

[0059] In order to improve the accuracy of similarity judgment, the high-level perceptual features φ(y) extracted by the VGG19 network are further introduced, where y is an image. VGG19 extracts high-level perceptual features of images through a pre-trained model, which enhances the understanding ability of similarity evaluation. The calculation formula for perceptual similarity is:

[0060]

[0061] Among them, φ(·) represents the features extracted by the VGG19 network at a specific layer, and y t is the target image, y best is the most similar reference image, and N is the feature dimension. Taking into account the cosine similarity and perceptual similarity, a comprehensive similarity score is defined:

[0062] sim total =sim(f t ,f best )+λ·L perceptual

[0063] Among them, λ is the weight coefficient (0.25 in this method), which is used to balance the contribution of the two similarity scores. In this way, it not only relies on the numerical similarity of the low-dimensional feature vector, but also makes full use of the high-level perceptual information extracted by VGG19 to ensure the comprehensiveness and accuracy of similarity judgment.

[0064] Step 4: Image restoration.

[0065] The goal of this step is to use the most similar image y best The information in , accurately completes the target image y t Specifically, the missing area appears as some x in the target image. j The y coordinates corresponding to j Values ​​are missing. The completion process involves locating the missing regions, extracting the corresponding data in the reference image, and applying smooth interpolation techniques.

[0066] First, by analyzing the target image y t The missing area in the target image is determined by determining the location of the missing points. Assume that the total number of coordinates in the target image is N, where there are some missing y j Value. Through the predefined mask matrix M∈{0,1} N to represent these missing positions. The matrix elements are defined as follows:

[0067]

[0068] On this basis, using similar images y best Extract its corresponding value, fill the missing data in the target image. For each missing position x j ∈M, the missing values ​​of the target image are filled with the corresponding values ​​in the most similar image. Although the missing values ​​can be effectively restored by directly filling in the similar images, this simple replacement may lead to unnatural transitions between the restored areas and the surrounding areas.

[0069] To solve this problem, the convolution kernel K is used to initially restore the image y interpolated Perform a smoothing operation to reduce the gradient difference between the restored area and the surrounding area to ensure visual continuity:

[0070] y smooth =K*y interpolated

[0071] Among them, * indicates that the convolution operation K is a smooth convolution kernel, y smooth is the restored image after smoothing. In order to further ensure the seamless transition between the restored area and the surrounding area, a global optimization method in image restoration, Poisson fusion technology, is also introduced. This optimization process can be achieved by minimizing the gradient difference in the restored area:

[0072]

[0073] in, Represents the gradient operation of the image, y recovered For the final restored image, is the restored data of the reference image.

[0074] Embodiment 1

[0075] Similar image search without missing images

[0076] In order to verify the hypothesis that there are similar patterns in a large amount of historical wind power data with future data, this experiment used millions of wind power data from dozens of wind farms at home and abroad. After processing, these time series data were converted into power curve images, and finally a database containing more than 100,000 images was constructed. The experiment was conducted on a high-performance server equipped with TeslaV100-32G*2, and the model construction time was about 120 minutes. For each power curve image in the test set, the model searched for similar images in the database, and the search time for a single image was about 50 seconds.

[0077] The experimental results are as follows Figure 4 As shown in the figure, the search results of multiple images in the test set in the database are demonstrated, and reference image segments that are highly similar to the test images in terms of macro trends and detailed features are successfully found. These results show that there are indeed patterns in historical data that are similar to future data, which verifies the effectiveness and reliability of the power data recovery method based on visual perception of the present invention.

[0078] Embodiment 2

[0079] Similar image search and data recovery of continuous multi-point missing images (image restoration level)

[0080] In practical applications, the missing power data of wind power generation often presents the characteristics of continuous multi-point missing, such as long-term data interruption caused by equipment failure or extreme weather. For the target power curve image with continuous multi-point missing, this method extracts its deep feature vector through pre-trained convolutional neural network combined with residual network and attention mechanism. These feature vectors are then compared with the features in the constructed database containing more than 100,000 historical power curve images. Through the comprehensive similarity scoring mechanism of cosine similarity and VGG19 perceptual features, it is possible to quickly and accurately find reference image segments that are highly similar to the target image in overall trends and detailed features.

[0081] After finding similar images, the corresponding data in the reference image segment is applied to the missing area of ​​the target image. At the same time, to ensure that the restored image is visually seamless, the smooth interpolation technology is used to process the initial restored image through the convolution kernel to reduce the gradient difference between the restored area and the surrounding area. Subsequently, the Poisson fusion technology is introduced to achieve seamless fusion of the restored area with the overall image by minimizing the gradient difference in the restored area.

[0082] Experimental results show that this method can effectively restore missing data and maintain data continuity and consistency when processing power curve images with continuous multi-point missing. Figure 5 As shown in the figure, after the image with continuous missing points is restored by this method, it is highly consistent with the original complete image in trend and details, with no obvious seams visually and significantly reduced numerical errors.

Claims

1. A power data recovery method based on visual perception, characterized in that: The steps include: Step 1, data collection and image generation: This step is responsible for collecting wind power data from multiple wind farms, dividing the data into time windows and converting them into power curve images to construct a data set; Step 2, feature extraction model training: In this step, the features of the wind power curve are extracted by combining the convolutional neural network autoencoder with the residual network and the attention mechanism. The feature extraction only involves the encoding stage of the autoencoder. The extracted features are saved as independent files for subsequent similarity calculation and data recovery tasks. Step 3, image similarity calculation and discrimination: This step calculates the similarity between image features based on a hybrid loss function combining the VGG19 algorithm and cosine similarity, compares the overall trend and fine features of the image, and finds the image segment that is most similar to the target image; Step 4, image restoration: This step further analyzes the missing area in the target image and determines the missing location. It combines the most similar image segments, performs smoothing through image restoration technology, interpolates the missing data, and ensures the continuity and accuracy of the restored data.

2. The power data recovery method based on visual perception according to claim 1 is characterized in that: In step 1, the wind power data is converted into power curve images in two ways: a fixed time window and a sliding time window. The former ensures the temporal continuity of the data, and the latter enhances the capture of different time feature patterns through diversified data input, and finally constructs a historical data image database containing 100,000 images.

3. The power data recovery method based on visual perception according to claim 1 is characterized in that: In step 2, a convolutional neural network is used in combination with a residual network and an attention mechanism to perform deep feature extraction on the power curve image. The residual network is used to improve the robustness of feature expression, and the attention mechanism focuses on key area features to generate a low-dimensional feature vector representation. At the same time, the generalization ability of the model is further optimized through a data enhancement method. During the data preprocessing process, the input wind power curve image is first uniformly adjusted to 224×224 pixels to meet the input requirements of the model. Then, a series of data enhancement techniques are applied, including random horizontal flipping, translation, scaling and rotation. These operations are implemented through a Python library to increase the diversity of training data and improve the generalization ability of the model. The normalization process is based on the mean and standard deviation commonly used in pre-training models, that is, the mean is (0.485, 0.456, 0.406) and the standard deviation is (0.229, 0.224, 0.225) to accelerate the convergence process of the model. The specific preprocessing and enhancement steps are as follows: x aug =Normalize(Transform(x)) Where x is the original input image, x aug is the image after data enhancement and normalization; The encoder part consists of multiple convolutional layers, batch normalization layers, ReLU activation functions, residual blocks, and attention mechanisms. The operation process of each convolutional layer is as follows: H (l) =RELU(BatchNorm(W (l) *H l-1 +b (l) )) Among them, H (l-1) is the input feature map of the l-1th layer, W (l) is the convolution kernel weight of the lth layer, b (l) is the bias term, ReLU is the activation function, * represents the convolution operation, H (l) is the output feature map of the lth layer; Subsequently, the residual block is introduced to enhance the feature extraction capability and alleviate the gradient vanishing problem in the deep network; the operation process of the residual block is as follows: Among them, x is the input feature map, W1 (l) and are the weight matrices of the two convolutional layers in the lth residual block, and is the corresponding bias vector, W s (l) The convolution kernel is used to adjust the number of channels to ensure the dimension matching of the jump connection. The jump connection directly adds the input feature map to the output after the convolution transformation to achieve the identity mapping and enhance the efficiency of information transmission. After each residual block, this method introduces an attention mechanism to further improve the effect of feature extraction; the attention mechanism calculates the importance weights of different areas in the feature map and dynamically adjusts the response strength of each channel of the feature map, so that the model can pay more attention to the key information in the image.

4. The power data recovery method based on visual perception according to claim 1 is characterized in that: In step 3, by calculating the cosine similarity between the target image and the feature vectors of all images in the database, multi-dimensional feature matching is performed in combination with the perceptual features of the VGG19 network, and the cosine similarity and perceptual feature scores are comprehensively used to accurately identify the reference image segments that are consistent with the target image in trend and similar in details; First, for the target image, the same preprocessing steps as in the training phase are applied, including resizing, data augmentation, and normalization, and then it is input into the encoder to generate the feature vector f of the target image. t , its mathematical expression is: f t =ε(x t ;θ ε ) Among them, x t is the target image, ε represents the encoder function, θ ε is the encoder parameter, f t is the feature vector of the target image; next, by calculating the feature vector of the target image f t With all image feature vectors f in the database i The cosine similarity between them quantifies the similarity between images. The calculation formula of cosine similarity is as follows: Among them, f t ·f i Represents the vector f t With f i The dot product of ||f t ||2 and ||f t ||2 are vectors f t and f i By calculating all similarity values, the image x that is closest to the target image in the feature space is identified. best , and its similarity score is: In order to improve the accuracy of similarity judgment, the high-level perceptual features φ(y) extracted by the VGG19 network are further introduced, where y is an image; VGG19 extracts high-level perceptual features of images through a pre-trained model, which enhances the understanding ability of similarity evaluation; the calculation formula of perceptual similarity is: Among them, φ(·) represents the features extracted by the VGG19 network at a specific layer, and y t is the target image, y best is the most similar reference image, N is the feature dimension; considering the cosine similarity and perceptual similarity, a comprehensive similarity score is defined: Yes total =yes(f t ,f best )+λ·L perceptual Among them, λ is the weight coefficient (taken as 0.25 in this method), which is used to balance the contribution of the two similarity scores; in this way, it not only relies on the numerical similarity of the low-dimensional feature vectors, but also makes full use of the high-level perceptual information extracted by VGG19 to ensure the comprehensiveness and accuracy of similarity judgment.

5. The power data recovery method based on visual perception according to claim 1 is characterized in that: In step 4, the missing data in the target image is supplemented by the most similar image segment, the smooth interpolation technology is used to optimize and repair the transition area, and the Poisson fusion technology is further used to reduce the gradient difference between the restored area and the surrounding area to ensure the continuity and visual consistency of the restored data; First, by analyzing the target image y t The missing area in the target image is determined to determine the location of the missing points. Let the total number of coordinates in the target image be N, where there are some missing y j Value; through the predefined mask matrix M∈{0,1} N to represent these missing positions; the matrix elements are defined as follows: On this basis, using similar images y best Extract its corresponding value, fill the missing data in the target image; for each missing position x j ∈M, the missing values ​​of the target image are filled with the corresponding values ​​in the most similar image; although the missing values ​​can be effectively restored by directly filling in the similar image, this simple replacement may lead to unnatural transitions between the restored area and the surrounding area; To solve this problem, the convolution kernel K is used to initially restore the image y interpolated Perform a smoothing operation to reduce the gradient difference between the restored area and the surrounding area to ensure visual continuity: and smooth =K*y interpolated Among them, * indicates that the convolution operation K is a smooth convolution kernel, y smooth is a smoothed restored image; in order to further ensure a seamless transition between the restored area and the surrounding area, a global optimization method in image restoration, Poisson fusion technology, is introduced; this optimization process can be achieved by minimizing the gradient difference in the restored area: in, Represents the gradient operation of the image, y recovered For the final restored image, is the restored data of the reference image.

6. The power data recovery method based on visual perception according to claim 1 is characterized in that: The sliding time window strategy is combined with the multi-stage image generation method to generate diversified power curve images to enhance the richness of the dataset. The model's accuracy in recovering missing data under different time, weather and equipment conditions is verified through cross-scenario testing.

Images