A time series image-based variable measurement data repair method

Abstract

The application discloses a kind of based on time series image variable measurement data repair method, it is related to power distribution network data management technical field, including obtaining distribution variable measurement multidimensional sequence, obtains normalized multidimensional sequence to distribution variable measurement multidimensional sequence normalization;Gram angle field method is used to process each dimension of normalized multidimensional sequence, multiple single-channel input feature map, the input feature map of each channel is spliced to obtain multi-channel input feature map;Different measurement missing strategy is designed, obtains the missing strategy used for model training after combination;Mask tensor is constructed to missing strategy, and mask tensor is multiplied by multi-channel input feature map to obtain the multi-channel input feature map after masking;Based on deep learning model, obtain multi-channel output feature map, and obtain repaired measurement multidimensional sequence based on the inverse operation of Gram angle field.The application can comprehensively consider distribution variable multidimensional measurement data characteristics, and realize end-to-end automatic repair for different measurement missing conditions.

Description

Technical Field

[0001] This invention relates to the field of distribution network data governance technology, specifically a method for repairing distribution network measurement data based on time-series visualization. Background Technology

[0002] As a crucial link connecting the transmission network and users, the accurate monitoring of the distribution network's operational status is essential for ensuring power supply reliability and optimizing energy allocation. Distribution transformers (DVTs), as core equipment in the distribution network, collect multi-dimensional measurement data such as current, voltage, active power, and reactive power in real time through deployed sensing terminals, comprehensively reflecting the equipment's operating status. However, due to factors such as communication link interference and sensor malfunctions, DVT measurement data often suffers from varying degrees of loss, making it difficult to guarantee the accuracy of load forecasting, fault diagnosis, and line loss analysis, adversely affecting the distribution network's scheduling decisions and operation and maintenance strategies. Therefore, constructing reliable measurement data repair methods to achieve high-precision restoration of missing data is a key technological requirement supporting the digital transformation of smart distribution networks.

[0003] Existing measurement data repair methods can be broadly categorized into two technical paths: statistical analysis-based and data-driven. Statistical analysis-based methods utilize simple mathematical statistical analysis to perform linear interpolation and historical mean filling for historical measurement data of distribution transformers, based on historical and current measurement data. These methods generally ignore the nonlinear trends of the measurement sequence and lack the ability to uncover the inherent patterns of time-series dependence of distribution transformer measurements. Data-driven methods use machine learning models such as recurrent neural networks to supplement missing measurements through prediction. While these methods consider the time-series dependence of measurements, they suffer from high computational complexity and limited generalization ability, especially when dealing with complex missing patterns (such as continuous multi-period missing data or multi-dimensional synchronous missing data). They fail to fully explore the correlations between multi-dimensional measurement data, and the repair results are insufficient to meet the actual needs of refined distribution network management. Summary of the Invention

[0004] To address the challenges of variable operating conditions and complex measurement loss patterns in distribution transformers, which make it difficult for traditional methods to achieve high-precision data repair, this invention proposes a time-series image-based method for repairing distribution transformer measurement data. This method is simple to calculate, has a clear principle, and can help distribution network professionals effectively correct measurement data with various loss patterns.

[0005] Technical solution: The present invention provides a method for repairing matching variable measurement data based on time series visualization, comprising the following steps:

[0006] S1. Obtain the multidimensional sequence of the matching variable and normalize the multidimensional sequence of the matching variable to obtain a normalized multidimensional sequence.

[0007] S2. Use the Gram angle field method to process each dimension of the normalized multidimensional sequence to generate multiple single-channel input feature maps, and then stitch all the single-channel input feature maps into a multi-channel input feature map.

[0008] S3. Design different measurement missing patterns and combine the measurement missing patterns to form a missing strategy for model training;

[0009] S4. Construct a mask tensor according to the missing strategy, perform a dot product between the mask tensor and the multi-channel input feature map to obtain a masked multi-channel input feature map, and concatenate the masked multi-channel input feature map with the mask tensor to obtain a multi-channel integrated input.

[0010] S5. Construct a convolutional neural network model, including an encoder, decoder, and discriminator;

[0011] S6. Model training: Input the multi-channel integrated input into the convolutional neural network model to obtain the multi-channel output feature map, and iteratively optimize the model parameters using a joint loss function that includes mask reconstruction loss and adversarial loss.

[0012] S7. Model Inference: Based on the actual measurement loss situation, a target multi-channel integrated input is constructed. The repaired multi-channel output feature map is output through the trained convolutional neural network model. Based on the inverse operation of the Gram angle field, the multi-channel output feature map is converted into a repaired multi-dimensional sequence of distribution transformer measurements. Here, the actual measurement loss situation refers to the actual missing values ​​of distribution transformer voltage, current, active power, and reactive power measurements.

[0013] Furthermore, S1 includes the following specific steps:

[0014] S11. Acquire data, including the three-phase voltage sequence U of the distribution transformer. A U B U C The three-phase current sequence I of the distribution transformer A I B I C The three-phase active power sequence P of the distribution transformer A P B P C The three-phase reactive power sequence Q of the distribution transformer A Q B Q C Where A, B, and C represent the A, B, and C phases of the distribution transformer voltage, respectively;

[0015] S12. Select the three-phase voltage sequences without missing values, and obtain the normalized voltage sequence U through bounded maximum-minimum normalization. ph norm ;

[0016]

[0017] Among them, U ph Represents a single-phase voltage sequence, U ub U represents the upper bound of the reasonable range of a single-phase voltage sequence. lb This represents the lower bound of the reasonable range for a single-phase voltage sequence.

[0018] S13. Filter out the three-phase current sequences without missing values, and obtain the normalized current sequence I through bounded maximum-minimum normalization. ph norm :

[0019]

[0020] Among them, I ph Represents a single-phase current sequence, I ub Represents the historical maximum value of a single-phase current sequence;

[0021] S14. Filter out the three-phase active power sequences without missing values, and obtain the normalized active power sequence P through bounded maximum-minimum normalization. ph norm :

[0022]

[0023] Among them, P ph Represents a single-phase active power sequence, P ub Represents the historical maximum value of a single-phase active power sequence;

[0024] S15. Filter out the three-phase reactive power sequence without missing values, and obtain the normalized reactive power sequence Q through bounded maximum-minimum normalization. ph norm :

[0025]

[0026] Among them, Q ph Represents a single-phase reactive power sequence, Q ub Represents the historical maximum value of a single-phase reactive power sequence;

[0027] S16. Concatenate the normalized voltage sequence U ph norm Current sequence I ph norm Active power sequence P ph norm and reactive power sequence Q ph norm Generate a normalized multidimensional sequence.

[0028] Furthermore, S2 includes the following specific steps:

[0029] S21. The sequence X = [x1,...,x] of each dimension of the normalized multidimensional sequence. l Transform into an angle sequence Φ=[φ1,...,φ] l ];

[0030]

[0031] Where l represents the length of the sequence;

[0032] S22. Construct a Gram angle field based on the angle sequence, and use it as a single-channel input feature map GASF:

[0033]

[0034] S23. Concatenate the input feature maps of each channel to obtain a multi-channel input feature map I. in .

[0035] Furthermore, the measurement missing modes include single-channel missing mode, multi-channel misaligned missing mode, multi-channel aligned missing mode, local missing mode, and large-scale missing mode.

[0036] Furthermore,

[0037] The single-channel missing mode includes introducing a missing feature map only for a single channel.

[0038] The multi-channel misaligned missing mode includes introducing missing data independently in different channels, so that the time steps of the missing parts are misaligned.

[0039] The multi-channel alignment missing mode includes introducing the same missing data in different channels to align the time steps of the missing parts.

[0040] The local deletion pattern includes introducing random deletion points or random deletion segments into the measurement sequence, with a total length not exceeding 20% ​​of the sequence length;

[0041] The large-scale deletion pattern includes introducing random deletion segments into the measurement sequence, with a total length of 20% to 80% of the sequence length.

[0042] Furthermore, S4 includes the following specific steps:

[0043] S41. Construct a mask tensor T for the missing strategy. mask Its shape is [feature map length, feature map width, number of channels], and the initial values ​​are all 1. According to the missing position in the missing strategy, the corresponding row and column of the mask tensor are set to 0;

[0044] S42. Calculate the Hadamard product between the mask tensor and the multi-channel input feature map to obtain the masked multi-channel input feature map I. mask The values ​​corresponding to the missing positions in the mask tensor are set to 0, which is represented as:

[0045]

[0046] S43. Concatenate the masked multi-channel input feature map with the mask tensor to obtain the multi-channel integrated input [I]. mask ;T mask Its shape is [feature map length, feature map width, number of channels × 2].

[0047] Furthermore, S5 includes the following specific steps:

[0048] S51. Construct an encoder f based on a deep learning model using a convolutional neural network to map the multi-channel integrated input to the latent space. This encoder includes convolutional layers, batch normalization layers, and ReLU activation layers to obtain the latent space feature map, represented as:

[0049]

[0050] Where θ represents the encoder parameters and z represents the latent space feature map;

[0051] S52. Construct a decoder g based on a deep learning model using a convolutional neural network to reconstruct the latent space feature map into an output feature map. This includes convolutional layers, deconvolutional layers, batch normalization layers, and ReLU activation layers to obtain a multi-channel output feature map with the same size as the input feature map and the same number of channels as the original feature map. This is represented as:

[0052]

[0053] in, The parameters representing the decoder, I out Represents the multi-channel output feature map;

[0054] S53. Construct a discriminator d based on a deep learning model using a convolutional neural network to determine whether the input image is the original image or an image reconstructed by the encoder and decoder. This discriminator includes convolutional layers, activation layers, and linear transformation layers, and obtains a scalar value, represented as:

[0055]

[0056] in, The parameters represent the discriminator, s represents a scalar, and I represents the input image, which may be a multi-channel input feature map I. in Or multi-channel output feature map I out .

[0057] Furthermore, S6 includes the following specific steps:

[0058] S61, including the following specific steps: using the encoder and decoder of a deep learning model to process the multi-channel integrated input, obtaining a multi-channel output feature map, assigning different weights to corresponding elements in the mask tensor, and calculating the mask reconstruction loss L. recon :

[0059]

[0060] Where i, j, and c represent the coordinates of the tensor in the length, width, and channel directions, respectively; Calculate weights for the reconstruction error of missing locations compared to normal locations. ;

[0061] S62. Use the discriminator of the deep learning model to process the multi-channel input feature map and the multi-channel output feature map, calculate the scalar distance between the output and the label indicating whether the input is the original image, and use it as the adversarial loss L. adv :

[0062]

[0063] in, For the Sigmoid function;

[0064] S63. Iteratively optimize the model using loss functions based on mask reconstruction loss and adversarial loss until the model loss stabilizes and no longer increases or stabilizes at a small value, that is, the model loss function converges, and then the model training is complete.

[0065] Furthermore, S7 includes the following specific steps:

[0066] S71. Construct a multi-channel input feature map based on the measurement data; the measurement data includes the voltage, current, active power, and reactive power time series data of the distribution transformer;

[0067] S72. Construct a mask tensor based on the actual missing data, and concatenate the multi-channel input feature map with the mask tensor to obtain a multi-channel integrated input.

[0068] S73. Based on the trained deep learning model, obtain multi-channel output feature maps;

[0069] S74. Based on the inverse operation of the Gram angle field, obtain the repaired multidimensional sequence of the coordinate variable. The data at time i is represented as:

[0070]

[0071] in For the multi-channel output feature map, it corresponds to the diagonal element in the i-th row and i-th column of the feature map.

[0072] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0073] (1) This invention adopts a data-driven approach, constructs a multi-channel input feature map based on a multivariate time series image method, and achieves comprehensive repair of multi-dimensional measurement data through a deep learning model. It can capture the correlation between multi-dimensional measurements and has strong accuracy.

[0074] (2) This invention generates a missing strategy by combining different missing modes, which effectively simulates complex measurement missing situations such as different channels, different ranges, alignment and misalignment in real operation, and overcomes the problem of low versatility of traditional methods.

[0075] (3) This invention uses multi-channel missing images and mask matrices as input, enabling the model to explicitly perceive the missing location. The calculation process is more targeted and interpretable, which can help power distribution professionals effectively correct measurement data of various missing patterns and has good application prospects. Attached Figure Description

[0076] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0077] Figure 1 This is a flowchart illustrating a method for repairing coordinate measurement data based on time series visualization, as disclosed in an embodiment of the present invention.

[0078] Figure 2 This is a comparison chart of the symmetrical mean absolute percentage error of the repair method of this invention and the methods based on linear interpolation and recurrent neural network prediction on the matching variable test sequences with different missing rates. Detailed Implementation

[0079] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable by the present invention will become clearer from the following detailed description.

[0080] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application, design, and conditions of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0081] In this invention, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0082] Example 1

[0083] Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for repairing coordinate measurement data based on time-series visualization, as disclosed in an embodiment of the present invention. Figure 1 The described method for repairing measurement data of distribution transformers is applicable to power systems, such as for repairing measurement data of public transformers, and is not limited to this application in the embodiments of the present invention. Figure 1 As shown, the matching variable measurement data repair method based on time series visualization can include the following operations:

[0084] S1. Obtain the multidimensional sequence of the matching variable and normalize it to obtain the normalized multidimensional sequence. The specific implementation process is as follows:

[0085] In the electricity information collection system (a system that exists in the prior art), select the distribution transformer and export the measurement data of the three-phase voltage, three-phase current, three-phase active power and three-phase reactive power of the distribution transformer with a sampling interval of 15 minutes. The aforementioned sampling interval of 15 minutes can be modified according to the actual situation.

[0086] After filtering out three-phase voltage sequences without missing values, a normalized voltage sequence U is obtained through bounded maximum-minimum normalization. ph norm :

[0087]

[0088] Among them, U ph Represents a single-phase voltage sequence, U ub U represents the upper bound of the reasonable range of a single-phase voltage sequence. lb This represents the lower bound of the reasonable range for a single-phase voltage sequence.

[0089] After filtering out the three-phase current sequences without missing values, the normalized current sequence I is obtained through bounded maximum-min normalization. ph norm :

[0090]

[0091] Among them, I ph Represents a single-phase current sequence, I ub Represents the historical maximum value of a single-phase current sequence;

[0092] By selecting the three-phase active power sequence without missing values, and then using bounded maximum-min normalization, the normalized active power sequence P is obtained. ph norm :

[0093]

[0094] Among them, P ph Represents a single-phase active power sequence, P ub Represents the historical maximum value of a single-phase active power sequence;

[0095] By selecting a three-phase reactive power sequence without missing values, and then performing bounded maximum-min normalization, a normalized reactive power sequence Q is obtained. ph norm :

[0096]

[0097] Among them, Q ph Represents a single-phase reactive power sequence, Q ub Represents the historical maximum value of a single-phase reactive power sequence;

[0098] splicing normalized voltage sequence U ph norm Current sequence I ph norm Active power sequence P ph norm Reactive power sequence Q ph norm , thus obtaining a normalized multidimensional sequence.

[0099] S2. The Gram angle field method is used to process each dimension of the normalized multidimensional sequence to obtain multiple single-channel input feature maps. These single-channel input feature maps are then concatenated to obtain a multi-channel input feature map. The specific implementation process is as follows:

[0100] The normalized multidimensional sequence is represented by the sequence X = [x1,...,x] for each dimension. l Transform into an angle sequence Φ=[φ1,...,φ]l ]:

[0101]

[0102] Where l represents the length of the sequence;

[0103] A Gram angle field is constructed based on the angle sequence and used as a single-channel input feature map GASF.

[0104]

[0105] The input feature maps of each channel are concatenated to obtain the multi-channel input feature map I. in .

[0106] This invention applies Gram angle field to the image encoding of multivariable distribution transformer time series data. By independently encoding the sequence of each dimension into a single-channel feature map and concatenating them into a multi-channel input, it achieves a visual representation of the temporal correlation between multiple dimensions. Compared with univariate GAF encoding, this invention can more effectively capture the physical correlation between multi-dimensional measurement sequences such as voltage, current, and active power.

[0107] S3. Design different measurement missing patterns and combine them to obtain a missing strategy for model training. The specific implementation process is as follows:

[0108] Designing a single-channel missing mode refers to introducing a missing feature into only a single channel in the feature map.

[0109] The design of a multi-channel misaligned missing mode refers to the independent introduction of missing data in different channels, resulting in misalignment of the time steps of the missing data.

[0110] Design a multi-channel aligned missing mode, which means introducing the same missing part into different channels so that the time steps of the missing part are aligned;

[0111] Designing a local deletion pattern refers to introducing random deletion points or random deletion segments into a measurement sequence, with the total length not exceeding 20% ​​of the sequence length;

[0112] Designing large-scale missing patterns refers to introducing random missing segments into a measurement sequence, with a total length greater than 20% but not greater than 80% of the sequence length;

[0113] Combining the above missing patterns yields single-channel local missing, single-channel large-scale missing, multi-channel aligned local missing, multi-channel misaligned local missing, multi-channel aligned large-scale missing, and multi-channel misaligned large-scale missing, which can be used as missing strategies for model training.

[0114] For example, for a single-channel local missing strategy, assuming the length of the variable measurement sequence is 96, one implementation is: for the A-phase voltage sequence, data is missing at time steps 10-11 and 50-54;

[0115] For example, for large-scale missing data in multi-channel alignment, assuming the length of the matching variable sequence is 96, one implementation is: for all dimensional sequences, data is missing at time steps 10-20, 50-70, and 80-85.

[0116] This invention designs five missing data modes: single-channel missing, multi-channel misaligned missing, multi-channel aligned missing, local missing, and large-scale missing. It also proposes a combined missing data strategy mechanism to simulate complex scenarios of missing data in real-world distribution transformer data. For example, the multi-channel aligned missing mode simulates the situation where sensor failure leads to synchronous missing of multi-dimensional measurements. In real distribution transformer measurement data, due to communication or sensor failures, there may be complex situations where all variables are collectively missing within the same time period, or where variables are randomly missing at different time periods. Existing single-variable repair methods or simple multi-variable missing data simulations cannot effectively cover such scenarios. The combined missing data construction strategy proposed in this invention considers the characteristics of multi-channel missing data and designs two differentiated strategies: multi-channel aligned and misaligned, ensuring that the trained model has the ability to cope with complex missing data situations in the real world.

[0117] S4. Construct a mask tensor for the missing strategy. Multiply the mask tensor by the multi-channel input feature map to obtain the masked multi-channel input feature map. Concatenate the mask tensor with the mask tensor to obtain the multi-channel integrated input. The specific implementation process is as follows:

[0118] Construct a mask tensor T for the missing strategy mask Its shape is [feature map length, feature map width, number of channels], and the initial values ​​are all 1. According to the missing position in the missing strategy, the corresponding row and column of the mask tensor are set to 0;

[0119] For example, for a single-channel local missing strategy, assuming the length of the variable measurement sequence is 96, the implementation is as follows: for the A-phase voltage sequence, if the data at time steps 10-11 and 50-54 is missing, then the data in the 10-11 row, 50-54 row, 10-11 column, and 50-54 column of the mask tensor in the feature map of the corresponding channel of the A-phase voltage are set to 0.

[0120] Calculate the Hadamard product between the mask tensor and the multi-channel input feature map to obtain the masked multi-channel input feature map I. mask The values ​​corresponding to the missing positions in the mask tensor are set to 0, which is represented as:

[0121]

[0122] The masked multi-channel input feature map is concatenated with the mask tensor to obtain the multi-channel synthesized input [I]. mask ;T mask Its shape is [feature map length, feature map width, number of channels × 2].

[0123] This invention not only uses a mask tensor to mask the feature map, but also innovatively concatenates the mask tensor with the masked feature map into a multi-channel integrated input, enabling the model to clearly perceive the missing location and improve the targeted repair of the missing region.

[0124] S5. Construct a deep learning model based on a convolutional neural network, including an encoder, decoder, and discriminator. The specific implementation process is as follows:

[0125] An encoder f, based on a convolutional neural network, is constructed to map the multi-channel integrated input to the latent space. This mapping includes, but is not limited to, combinations of convolutional layers, batch normalization layers, and ReLU activation layers, resulting in a latent space feature map with a smaller feature map size and more channels. This is represented as:

[0126]

[0127] in, The parameters represent the encoder, and z represents the latent space feature map;

[0128] For example, an encoder implementation uses two convolution kernel parameters: a 3×3 kernel with a stride of 1 and padding of 1, and a 2×2 kernel with a stride of 2 and padding of 0.

[0129] A decoder g based on a deep learning model using a convolutional neural network is constructed to reconstruct the latent space feature map into an output feature map. This reconstruction includes, but is not limited to, combinations of convolutional layers, deconvolutional layers, batch normalization layers, and ReLU activation layers. The result is a multi-channel output feature map with the same size as the input feature map and the same number of channels as the original feature map, represented as:

[0130]

[0131] in, The parameters representing the decoder, I out Represents the multi-channel output feature map;

[0132] For example, one implementation of a decoder uses two convolutional kernel parameters: a 3×3 kernel with a stride of 1 and padding of 1, and a 2×2 deconvolutional kernel with a stride of 2 and padding of 0.

[0133] Construct a discriminator d based on a deep learning model using a convolutional neural network to determine whether the input image is the original image or a reconstructed image from the encoder and decoder. This discriminator includes, but is not limited to, combinations of convolutional layers, activation layers, and linear transformation layers. A scalar value is obtained, represented as:

[0134]

[0135] in, The parameters represent the discriminator, s represents a scalar, and I represents the input image, which may be a multi-channel input feature map I. in Or multi-channel output feature map I out .

[0136] For example, one implementation of the discriminator uses a convolution kernel with one set of parameters: a 3×3 kernel with a stride of 2 and padding of 1.

[0137] For multi-channel image input, this invention designs a multi-scale convolutional kernel combination to enhance the ability to distinguish multi-dimensional features. The decoder adopts a deconvolution and skip connection structure to avoid loss of detail.

[0138] S6. In the model training phase, a deep learning model is used to process the multi-channel integrated input to obtain multi-channel output feature maps. The model is then iteratively optimized using loss functions based on mask reconstruction loss and adversarial loss. The specific implementation process is as follows:

[0139] The encoder and decoder of a deep learning model process the multi-channel integrated input to obtain a multi-channel output feature map. Different weights are assigned to corresponding elements in the mask tensor, and the mask reconstruction loss L is calculated. recon :

[0140]

[0141] Where i, j, and c represent the coordinates of the tensor in the length, width, and channel directions, respectively; T mask (i,j,c) represents a specific element in the mask tensor where i is the length, j is the width, and c is the channel. in (i,j,c) represents a specific element in the multi-channel input feature map with length i, width j, and channel c; out (i,j,c) represents a specific element in the multi-channel output feature map with i in the length direction, j in the width direction, and c in the channel direction; Calculate weights for the reconstruction error of missing locations compared to normal locations. ;

[0142] For example, in order to improve the model's ability to reconstruct missing parts during training, the following methods are employed: ;

[0143] The discriminator of the deep learning model processes the multi-channel input feature map and the multi-channel output feature map, and calculates the scalar distance between the output and the label indicating whether the input is the original image, as the adversarial loss L. adv :

[0144]

[0145] in, For the Sigmoid function;

[0146] The model is iteratively optimized using loss functions based on mask reconstruction loss and adversarial loss until the model loss stabilizes and no longer increases, or stabilizes at a small value, that is, the model loss function converges, and then the model training is complete.

[0147] This invention proposes a weighted mask reconstruction loss function that assigns higher weights to missing positions. This forces the model to prioritize learning to reconstruct the missing parts. This design significantly improves repair accuracy, especially in complex scenarios such as multi-channel misalignment and missing data, where the repair error is reduced.

[0148] S7. In the model inference stage, a multi-channel integrated input is constructed based on the actual missing measurement information. A multi-channel output feature map is obtained based on a deep learning model, and the repaired multi-dimensional measurement sequence is obtained based on the inverse operation of the Gram angle field. The specific implementation process is as follows:

[0149] Based on the measurement data, construct a multi-channel input feature map;

[0150] Based on the actual missing data measured, a mask tensor is constructed, and the multi-channel input feature map and the mask tensor are concatenated to obtain a multi-channel integrated input.

[0151] Based on the trained deep learning model, multi-channel output feature maps are obtained;

[0152] Based on the inverse operation of the Gram angle field, the repaired measurement multidimensional sequence is obtained. The data at time i is represented as:

[0153]

[0154] in For the multi-channel output feature map, it corresponds to the diagonal element in the i-th row and i-th column of the feature map.

[0155] Those skilled in the art would find it difficult to readily deduce multivariate, multi-missing-mode repair schemes from univariate GAF repair methods, especially for handling multi-channel misaligned missing data, which requires redesigning the masking mechanism and loss function—something not achievable with conventional techniques. For example, a comparative experiment was conducted using the repair method of this invention with methods based on linear interpolation and recurrent neural network prediction on matched variable test sequences with different missing rates, using Symmetric Mean Absolute Percentage Error (SMAPE) as the metric. The results are as follows: Figure 2 As shown, under different measurement missing rates, the repair performance of this invention is superior to methods based on linear interpolation and recurrent neural network prediction.

[0156] In multi-channel misalignment missing scenarios, the present invention reduces SMAPE error by approximately 15%; even with a high missing rate of 80%, the error remains below 5%.

[0157] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for repairing matching variable measurement data based on time series visualization, characterized in that, Includes the following steps: S1. Obtain the multidimensional sequence of the matching variable and normalize the multidimensional sequence of the matching variable to obtain a normalized multidimensional sequence. S2. Use the Gram angle field method to process each dimension of the normalized multidimensional sequence to generate multiple single-channel input feature maps, and then stitch all the single-channel input feature maps into a multi-channel input feature map. S3. Design different measurement missing patterns and combine the measurement missing patterns to form a missing strategy for model training; S4. Construct a mask tensor according to the missing strategy, perform a dot product between the mask tensor and the multi-channel input feature map to obtain a masked multi-channel input feature map, and concatenate the masked multi-channel input feature map with the mask tensor to obtain a multi-channel integrated input. S5. Construct a convolutional neural network model, including an encoder, decoder, and discriminator; S6. Model training: Input the multi-channel integrated input into the convolutional neural network model to obtain the multi-channel output feature map, and iteratively optimize the model parameters using a joint loss function that includes mask reconstruction loss and adversarial loss. S7. Model inference: Construct a target multi-channel integrated input based on the actual measurement loss situation, output the repaired multi-channel output feature map through the trained convolutional neural network model, and convert the multi-channel output feature map into the repaired coordinate variable measurement multidimensional sequence based on the inverse operation of the Gram angle field. S1 includes the following steps: S11. Acquire data, including the three-phase voltage sequence of the distribution transformer. U A , U B , U C Three-phase current sequence of the distribution transformer I A , I B , I C Three-phase active power sequence of the distribution transformer P A , P B , P C Three-phase reactive power sequence of the distribution transformer Q A , Q B , Q C Where A, B, and C represent the A, B, and C phases of the distribution transformer voltage, respectively; The measurement missing modes include single-channel missing mode, multi-channel misaligned missing mode, multi-channel aligned missing mode, local missing mode, and large-scale missing mode; The single-channel missing mode includes introducing a missing feature map only for a single channel. The multi-channel misaligned missing mode includes introducing missing data independently in different channels, so that the time steps of the missing parts are misaligned. The multi-channel alignment missing mode includes introducing the same missing data in different channels to align the time steps of the missing parts. The local deletion pattern includes introducing random deletion points or random deletion segments into the measurement sequence, with a total length not exceeding 20% ​​of the sequence length; The large-scale deletion pattern includes introducing random deletion segments into the measurement sequence, with a total length of 20% to 80% of the sequence length; S4 includes the following specific steps: S41. Construct a mask tensor for the missing strategy. T mask Its shape is [feature map length, feature map width, number of channels], and the initial values ​​are all 1. According to the missing position in the missing strategy, the corresponding row and column of the mask tensor are set to 0; S42. Calculate the Hadamard product between the mask tensor and the multi-channel input feature map to obtain the masked multi-channel input feature map. I mask The values ​​corresponding to the missing positions in the mask tensor are set to 0, which is represented as: in, I in For multi-channel input feature maps; S43. Concatenate the masked multi-channel input feature map with the mask tensor to obtain the multi-channel integrated input. I mask ; T mask Its shape is [feature map length, feature map width, number of channels × 2].

2. The method for repairing coordinate measurement data based on time series visualization according to claim 1, characterized in that: S1 also includes the specific steps following step S11: S12. Filter out the three-phase voltage sequences without missing values, and obtain the normalized voltage sequences through bounded maximum-minimum normalization. U ph norm ; in, U ph Represents a single-phase voltage sequence. U ub This represents the upper bound of the reasonable range for a single-phase voltage sequence. U lb This represents the lower bound of the reasonable range for a single-phase voltage sequence. S13. Filter out the three-phase current sequences without missing values, and obtain the normalized current sequence through bounded maximum-min normalization. I ph norm : in, I ph Represents a single-phase current sequence. I ub Represents the historical maximum value of a single-phase current sequence; S14. Filter out the three-phase active power sequences without missing values, and obtain the normalized active power sequence through bounded maximum-min normalization. P ph norm : in, P ph Represents a single-phase active power sequence. P ub Represents the historical maximum value of a single-phase active power sequence; S15. Filter out the three-phase reactive power sequences without missing values, and obtain the normalized reactive power sequence through bounded maximum-min normalization. Q ph norm : in, Q ph Represents a single-phase reactive power sequence. Q ub Represents the historical maximum value of a single-phase reactive power sequence; S16. Concatenate the normalized voltage sequence U ph norm Current sequence I ph norm Active power sequence P ph norm and reactive power sequence Q ph norm Generate a normalized multidimensional sequence.

3. The method for repairing coordinate measurement data based on time series visualization according to claim 1, characterized in that: S2 includes the following specific steps: S21. The sequence of each dimension of the normalized multidimensional sequence. X =[ x 1,..., x l Convert to angle sequence Φ =[ Φ 1,..., Φ l ]; in, l Represents the length of the sequence; S22. Construct a Gram angle field based on the angle sequence, and use it as a single-channel input feature map GASF: S23. Concatenate the input feature maps of each channel to obtain a multi-channel input feature map. I in .

4. The method for repairing matching variable measurement data based on time series visualization according to claim 1, characterized in that: S5 includes the following specific steps: S51. Constructing an encoder for a deep learning model based on a convolutional neural network. f This is used to map the multi-channel synthesized input to the latent space, including convolutional layers, batch normalization layers, and ReLU activation layers, to obtain the latent space feature map, represented as: in, θ The parameters representing the encoder, z Represents the potential spatial feature map; S52. Constructing a decoder for a deep learning model based on a convolutional neural network. g This is used to reconstruct the latent space feature map into an output feature map, including convolutional layers, deconvolutional layers, batch normalization layers, and ReLU activation layers, to obtain a multi-channel output feature map with the same size as the input feature map and the same number of channels as the original feature map, represented as: in, Parameters representing the decoder, I out Represents the multi-channel output feature map; S53. Constructing a discriminator for a deep learning model based on a convolutional neural network. d It is used to determine whether the input image is the original image or the image reconstructed by the encoder and decoder. It includes convolutional layers, activation layers, and linear transformation layers, and obtains a scalar value, represented as: in, The parameters representing the discriminator, s Represents a scalar, I The input image represents a multi-channel input feature map. I in Or multi-channel output feature map I out .

5. The method for repairing matching variable measurement data based on time series visualization according to claim 4, characterized in that: S6 includes the following specific steps: S61, including the following specific steps: using the encoder and decoder of a deep learning model to process the multi-channel integrated input, obtaining a multi-channel output feature map, assigning different weights to corresponding elements in the mask tensor, and calculating the mask reconstruction loss. L recon : Where i, j, and c represent the coordinates of the tensor in the length, width, and channel directions, respectively; Calculate weights for the reconstruction error of missing locations compared to normal locations. ; S62. Use the discriminator of a deep learning model to process multi-channel input feature maps and multi-channel output feature maps, and calculate adversarial loss. L adv The calculation formula is as follows: in, For the Sigmoid function; S63. Iteratively optimize the model using loss functions based on mask reconstruction loss and adversarial loss until the model loss stabilizes and no longer increases, then the model training is complete.

6. The method for repairing coordinate measurement data based on time series visualization according to claim 1, characterized in that: S7 includes the following specific steps: S71. Construct a multi-channel input feature map based on the measurement data; S72. Construct a mask tensor based on the actual missing data, and concatenate the multi-channel input feature map with the mask tensor to obtain a multi-channel integrated input. S73. Based on the trained deep learning model, obtain multi-channel output feature maps; S74. Based on the inverse operation of the Gram angle field, obtain the repaired multidimensional sequence of the coordinate variable. ,in i The data at time point is represented as follows: in In the multi-channel output feature map, the corresponding feature map number is... i Line 1 i The diagonal elements of the column.

Images