A method, system, storage medium and device for predicting oil temperature of power transformer

Abstract

The present invention provides a power transformer oil temperature prediction method, system, storage medium, and device. These methods preprocess acquired power transformer oil temperature data; decompose the preprocessed data into high-frequency and low-frequency components; improve the neural network model by using a segmented attention mechanism instead of a dot-product attention mechanism. The improved neural network model extracts global and local features from the low-frequency component and projects the extracted features onto a prediction length; predicts the high-frequency component of the data using a linear model; and fuses the two prediction results to obtain a final prediction result. This method reduces data noise, fully interprets the data, and improves the prediction accuracy of complex and variable power transformer oil temperature data.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer oil temperature prediction, and in particular relates to a method, system, storage medium and equipment for predicting the oil temperature of a power transformer. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Power transformers are one of the important electrical equipment in the power system. The oil temperature of the transformer can effectively reflect the working condition of the power transformer. Predicting the oil temperature of the transformer and formulating power distribution and transformer maintenance strategies based on the prediction results can effectively ensure the safe operation of the transformer and help extend the service life of the transformer. It can also distribute power to different user areas according to the future needs of each user area, rationally plan power production, and avoid unnecessary waste.

[0004] However, predicting future transformer oil temperature data is difficult because it varies with factors such as weekdays, holidays, seasons, weather, and temperature. Transformer oil temperature prediction methods based on traditional time series models and machine learning methods are not suitable for high-precision, long-term predictions of real-world data. With the continuous development of deep learning, power transformer oil temperature prediction methods based on the Transformer model can better explore the patterns and characteristics in power transformer data, improving prediction precision and accuracy.

[0005] However, according to the inventors’ understanding, some current transformer oil temperature prediction methods based on the Transformer model still have the following problems:

[0006] The collected power transformer data is usually interfered by noise, which affects the accuracy of the prediction model, but the existing power transformer oil temperature prediction models often do not include effective denoising.

[0007] If the original sequence is used directly to predict the oil temperature of power transformers, the internal characteristics of the time series data will be mixed. Traditional time series prediction methods usually use a simple sliding average method to decompose the data. This decomposition method is often insufficient and has low interpretability.

[0008] The global self-attention mechanism in the Transformer model is not sensitive enough to local semantics when processing time series data, which leads to a decrease in the performance of the model in capturing dependencies in time series data.

[0009] In addition, for long-term time series prediction, the computational cost of the traditional prediction model based on the Transformer attention mechanism is also high. Summary of the Invention

[0010] In order to solve the above problems, the present invention proposes a method, system, storage medium and device for predicting the oil temperature of a power transformer. The present invention preprocesses the data and decomposes it into high-frequency and low-frequency components. The two parts of data are processed and predicted separately, and the two parts of prediction results are combined as the final prediction result of the model. This reduces data noise, fully explains the data, and improves the prediction accuracy of complex and changeable power transformer oil temperature data.

[0011] According to some embodiments, the present invention adopts the following technical solutions:

[0012] A method for predicting oil temperature of a power transformer comprises the following steps:

[0013] Preprocessing the acquired power transformer oil temperature data;

[0014] Decomposing the pre-processed power transformer oil temperature data to obtain high-frequency component and low-frequency component;

[0015] The segmented attention mechanism is used instead of the dot product attention mechanism to improve the neural network model. The improved neural network model is used to extract the global and local features of the low-frequency component, and the extracted features are projected to the prediction length to obtain the first prediction result.

[0016] Using a linear model to predict the high-frequency component of the data, a second prediction result is obtained;

[0017] The first prediction result and the second prediction result are integrated to obtain the final prediction result.

[0018] In the above technical solution, the internal features of the power transformer oil temperature data are separated, and the convolution-enhanced segmented attention mechanism is used to replace the traditional dot product attention mechanism in the model to model the low-frequency component of the transformer oil temperature data. While reducing the quadratic complexity of the dot product attention mechanism, local features are introduced to enhance the model's representation ability of the local semantics of time series data. The prediction results of the high-frequency and low-frequency parts are integrated, which can not only make the power transformer oil temperature data more fully decomposed, but also make independent predictions on the decomposed data in a targeted manner, ensuring the accuracy, timeliness and economy of the prediction of each part. In addition, the integration of the two parts of the prediction results improves the accuracy of the overall prediction results.

[0019] As an optional implementation, the preprocessing process includes normalization and denoising, normalizing the acquired power transformer oil temperature data, and using Fourier transform to convert the normalized time series from the time domain representation to the frequency domain, and performing denoising operation.

[0020] The data is preprocessed, and considering that the characteristics and joint distribution of non-stationary data change over time, the preprocessing process includes normalization and denoising, which can reduce the non-stationarity of the time series, thereby achieving better predictability and improving processing efficiency and effectiveness.

[0021] As a further step, the frequency band representing the noise is removed in the frequency domain, the first K dimensions are retained, and the remaining dimensions are randomly sampled, and s components are selected from the L Fourier components for storage, where K, L, and s are all integers, and s <L。

[0022] The above scheme can achieve the purpose of removing noise and prevent overfitting of historical data.

[0023] As a further step, the denoised data sequence is completed and padded, and then an inverse Fourier transform is performed to convert the frequency domain data back to the time domain wave.

[0024] As an optional implementation method, the process of decomposing the preprocessed power transformer oil temperature data includes decomposing the preprocessed power transformer oil temperature data using a variational modal decomposition method to obtain multiple modal components and residual terms, wherein the modal components are used as the low-frequency component part of the power transformer oil temperature data, and the residual terms are used as the high-frequency component part of the power transformer oil temperature data.

[0025] As an optional implementation, the specific process of extracting the global features and local features of the low-frequency component using the improved neural network model includes:

[0026] The low-frequency component is segmented according to the set segment length and division step, and the modal component is divided into segment structures;

[0027] The input of the segment structure is model embedded and position encoded;

[0028] The low-frequency component of the power transformer oil temperature data is modeled through the Transformer model, and the features extracted from the low-frequency component of the power transformer oil temperature data are obtained.

[0029] As a further implementation method, the modeling process of the low-frequency component of the power transformer oil temperature data using the Transformer model includes:

[0030] Generate a linear projection query matrix, a key matrix and a value matrix about the input through linear projection;

[0031] Split the input query matrix, key matrix, and value matrix according to the principle of multi-head attention;

[0032] Calculate the attention distribution of each segment, concatenate the output results of all heads as the output of the multi-head attention, and then pass it through the normalization layer and the feedforward network layer with residual connection to obtain the intermediate modeling result;

[0033] A convolutional structure is introduced in the hidden dimension to help capture local information within each segment, and the global and local features of the data are found through the global information network and the local information network respectively.

[0034] The above technical solution replaces the traditional dot-product attention mechanism in the Transformer model with a convolution-enhanced segmented attention mechanism to model the low-frequency component of the transformer oil temperature data. While reducing the quadratic complexity of the dot-product attention mechanism, it introduces local dependency modeling to enhance the model's ability to represent the local semantics of time series data.

[0035] As an optional implementation, the specific process of fusing the first prediction result and the second prediction result includes flattening the first prediction result and directly projecting it to a specified dimension, adding it to the second prediction result, and performing data denormalization on the added prediction result to obtain the final prediction result.

[0036] A power transformer oil temperature prediction system, comprising:

[0037] a preprocessing module configured to preprocess the acquired power transformer oil temperature data;

[0038] a decomposition module configured to decompose the preprocessed power transformer oil temperature data to obtain a high-frequency component part and a low-frequency component part;

[0039] A low-frequency component feature extraction module is configured to use a segmented attention mechanism instead of a dot-product attention mechanism to improve the neural network model, extract global and local features of the low-frequency component using the improved neural network model, and project the extracted features onto a prediction length to obtain a first prediction result;

[0040] a high-frequency component feature extraction module configured to predict the high-frequency component of the data using a linear model to obtain a second prediction result;

[0041] The prediction module is configured to fuse the first prediction result and the second prediction result to obtain a final prediction result.

[0042] As an optional implementation, the preprocessing module includes a data normalization module and a denoising module. The data normalization module is configured to perform normalization processing on the acquired power transformer oil temperature data, and the denoising module is configured to use Fourier transform to convert the normalized time series from a time domain representation to a frequency domain to perform denoising operation.

[0043] As a further step, the denoising module includes:

[0044] A Fourier transform module is configured to transform the normalized time series from a representation in the time domain to a frequency domain;

[0045] The noise removal module is configured to remove the frequency band representing the noise in the frequency domain, retain the first K dimensions, and randomly sample the remaining dimensions, selecting s components from the L Fourier components for storage, where K, L, and s are all integers, and s <L;

[0046] A completion module, configured to complete and fill the output of the noise removal module;

[0047] The inverse Fourier transform module is configured to perform inverse Fourier transform to convert the frequency domain data back into a time domain wave.

[0048] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps in the above method are completed.

[0049] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the above method are completed.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention utilizes the time-frequency domain changes of time series data to perform denoising operations on the oil temperature data of power transformers; utilizes the variational mode decomposition method to make the oil temperature data of power transformers more fully decomposed and separate the internal features of the oil temperature data of power transformers; then uses the improved convolution enhanced segmented attention mechanism to model the low-frequency component of the oil temperature of power transformers, uses a linear layer to project the extracted features to the prediction length, and directly uses a fully connected network to predict the high-frequency component of the data. The two prediction results are combined as the final prediction result of the model, thereby improving the prediction accuracy of the model for complex and changeable oil temperature data of power transformers.

[0052] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0054] Figure 1A flowchart of a method for predicting oil temperature of a power transformer based on an improved attention mechanism proposed in an embodiment;

[0055] Figure 2 A structural diagram of a Fourier denoising module proposed in one embodiment;

[0056] Figure 3 This is a structural diagram of the convolution-enhanced segmented attention mechanism in the low-frequency component feature extraction module for time series data proposed in one embodiment;

[0057] Figure 4 A structural diagram of a high-frequency component prediction module for time series data and a linear layer in the prediction module proposed in one embodiment. DETAILED DESCRIPTION

[0058] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0059] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0060] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0061] In the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0062] Example 1

[0063] A method for predicting oil temperature data of a power transformer comprises the following steps:

[0064] (1) Normalize the time series instance X with the mean and standard deviation to help alleviate the distribution shift effect between training data and test data, reduce the non-stationarity of the original time series, thereby achieving better predictability and providing a more stable data distribution for time series deep learning models.

[0065] Of course, it is necessary to introduce a denormalization operation in the data output layer to return the information deleted by the normalization during input to the model.

[0066] (2) Use Fourier transform to convert time series data from time domain representation to frequency domain representation, remove the band representing noise in the frequency domain, and then convert it back to time domain representation to achieve the purpose of removing noise.

[0067] (3) The variational mode decomposition (VMD) method is used to decompose the oil temperature data of the power transformer. The original oil temperature data of the power transformer is decomposed into a high-frequency component and a low-frequency component, and the two parts of data are independently modeled and processed.

[0068] (4) The convolution-enhanced segmented attention mechanism is used to replace the traditional dot product attention mechanism in the Transformer model to model the low-frequency component of the transformer oil temperature data. While reducing the quadratic complexity of the dot product attention mechanism, local dependency modeling is introduced to enhance the model's ability to represent the local semantics of time series data.

[0069] (5) A simple linear regression strategy is used to predict the high-frequency component of the power transformer oil temperature data.

[0070] (6) The low-frequency component features of the data are input into the linear layer to predict the future oil temperature of the power transformer, and added to the high-frequency component prediction results as the final prediction result.

[0071] Example 2

[0072] A power transformer oil temperature data prediction system, for the convenience of understanding of those skilled in the art, the power transformer oil temperature prediction problem is defined as a time series sample set given a lookback window L Predict the value T moments in the future

[0073] like Figure 1 As shown, including:

[0074] (1) Normalization module and denormalization module

[0075] The characteristics of non-stationary data are that the statistical properties of the data (such as mean, variance, etc.) and the joint distribution change over time. This characteristic will damage the predictability of the time series, so it is necessary to introduce a normalization module to normalize the time series sample set. Perform stationary preprocessing to generate a sequence Will The input is then fed into the subsequent model for modeling to reduce the non-stationarity of the time series, thereby achieving better predictability and providing a more stable data distribution for the deep learning model. The normalization module is shown in formula (1).

[0076]

[0077] ∈ is a numerical stability factor, and its default value is usually set to 1e-5, E k [X] and Var[X] represent the mean and standard deviation of the time series, respectively, and are defined as shown in Equations (2) and (3).

[0078]

[0079]

[0080] The subsequent model accepts the transformed data as input. The normalization module reduces the distribution differences of each input time series, making the model input more stable. The model predicts future values based on the input and explicitly returns the non-stationary attributes removed from the input data to the model output by performing scaling and shifting on the model output at the symmetric position, i.e., the inverse normalization step of the model output layer, using the same statistics as those used in the normalization step of Equation (2) to generate the final model input Y, as shown in Equation (4). Where E is scaled and shifted, generating the final model input Y, as shown in Equation (4). Where E t [X] and Var[X] are defined the same as in Equations (2) and (3).

[0081]

[0082] (2) Denoising module

[0083] Time series data can be modeled from two aspects: the time domain and the frequency domain. Fourier analysis is a common tool for delving into the frequency domain. By using Fourier operations, a time series can be transformed from its representation in the time domain to the frequency domain. The normalized oil temperature data of the power transformer is subjected to Fourier transform, resulting in The Fourier transform is shown in Equation (5), where is the given real sequence in the time domain, n = 0,​​​​​​​​, then S i,k =1, otherwise S i,k = 0, then the denoising operation can be expressed as shown in formula (6). After the denoising operation is completed, the sequence After the padding operation, it is filled Then, the inverse Fourier transform is used to convert the frequency domain data back to the time domain wave. The inverse Fourier transform is shown in formula (7), where k represents the index of different frequency components in the frequency domain, and t represents the time point to be calculated when we convert the frequency domain data back to the time domain signal. The denoising module structure based on Fourier transform is as follows: Figure 2 shown.

[0085]

[0086]

[0087]

[0088] (3) VMD decomposition module

[0089] VMD can decompose the original sequence data into multiple components with specific bandwidths on the spectrum. The prediction model based on VMD decomposition has higher prediction accuracy than the models based on other decomposition methods. After VMD decomposition, a smoother subsequence can be obtained, thus more effective modeling.

[0090] Given the processed power transformer oil temperature data Use VMD decomposition algorithm to decompose it into K modal components μ k (t) and the residual term r(t), the VMD decomposition is shown in formula (8).

[0091]

[0092] The main idea of ​​VMD is to obtain the modal components μ by optimizing the following variational problem, as shown in formula (9): k (t) and the residual term r(t):

[0093]

[0094] Among them, u k (t) is the kth modal component, ω k is the corresponding frequency, represents the Hilbert transform, and α is the regularization parameter used to balance fitting and smoothness.

[0095] The following are the specific implementation steps for using the VMD algorithm to decompose power transformer oil temperature data:

[0096] a. Initialize VMD parameters, including the number of modes K and the regularization parameter α.

[0097] b. Initialize a set of initial modal components u k (t), usually generated using random numbers.

[0098] c. For each modal component u k (t) Perform iterative optimization to satisfy the optimization problem of formula (9).

[0099] d. Iterate the above optimization problem until the modal components converge. This will produce K modal components u k (t), where each modal component corresponds to a different frequency and amplitude.

[0100] The K modal components are used as the low-frequency component part of the power transformer oil temperature data, and the residual term is used as the high-frequency component part of the power transformer oil temperature data and input into the subsequent modules for modeling respectively.

[0101] (4) Low-frequency component feature extraction module

[0102] A segmented attention mechanism is used instead of the traditional dot-product attention mechanism to model the low-frequency components of power transformer oil temperature data. Transformers have demonstrated powerful long-sequence feature extraction capabilities in natural language processing. Through multi-head attention and residual concatenation, they can compute the entire sequence and extract global features at once. However, the quadratic computational complexity of the Transformer model poses significant challenges in training and inference efficiency when applied to large-scale real-world datasets. Therefore, the decomposed low-frequency components are first segmented, and then a self-attention mechanism is applied to each segment to reduce computational complexity.

[0103] The low-frequency component is modeled using the convolution-enhanced segmented attention mechanism. Specifically:

[0104] For k modal components Perform segmentation operation, the segment length is S, the step size of segment division is S / 2, and the input modal component is divided into the input of the segment structure Where N is the number of segments divided,

[0105] Then, the input of the segment structure is model embedded and position encoded, as shown in formula (10), is a trainable linear layer used to project the segmented structure into the model dimension, PE represents the position encoding matrix, Positional encoding is represented by sine-cosine positional encoding, as shown in formulas (11) and (12), where pos represents the position in the segment structure input, i represents the dimension in the positional encoding, and D represents the embedding dimension of the model.

[0106] Xemb =X input W mlp +PE (10)

[0107]

[0108]

[0109] Then, the low-frequency component of the power transformer oil temperature data is first modeled through Transformer. The modeling process is shown in Equations (13) to (20).

[0110] 1. Generate about input X through linear projection emb The linear projection query matrix Q, key matrix K, value matrix V, where F q ,F k ,F v Both are linear layers:

[0111] Q=F q (X emb ),K=F k (X emb ),V=F v (X emb ) (13)

[0112] 2. Split the input Q, K, V according to the principle of multi-head attention:

[0113] [q1,q2,…,q H ]=MultiHead(Q) (14)

[0114] [(k1,c),(k2,v2),…,(k H ,v H )]=MultiHead(K,V) (15)

[0115] 3. Calculate the attention distribution of each segment, concatenate the output results of all heads as the output of multi-head attention, and then pass it through the Norm layer and the feedforward network layer with residual connection to obtain the modeling intermediate result

[0116]

[0117] M i =Softmax(E i )v i (17)

[0118] Attn=Contact(M1, M2,...,M H ) (18)

[0119] Z=Norm(Attn)+X emb (19)

[0120] Z′=Norm(FFN(Z))+Z (20)

[0121] The Transformer model relies on the self-attention mechanism to calculate the weights of data segments to determine the long-term dependencies of the sequence, but the attention mechanism often focuses mainly on global dependencies and pays less attention to local information.

[0122] In the above time series segmentation, local information reflects the time-varying pattern of the data, such as Figure 3 As shown, an additional convolutional structure is introduced in the hidden dimension to help capture local information within each segment. By finding the global and local features of the data through the global information network and local information network, respectively, the completeness and effectiveness of modeling the long-term characteristics of power transformer oil temperature data are greatly improved.

[0123] Formula (21)-Formula (25) is the process of introducing additional convolution structure to model the oil temperature data of power transformer, Z Low The features extracted from the low-frequency component of the final power transformer oil temperature data.

[0124] Z1=LayerNorm(Z) (21)

[0125] Z2=Gelu(PointWiseConv(Z1)) (22)

[0126] Z3=DepthWiseConv(Z2) (23)

[0127] Z4=LeakyRelu(Norm(Z3)) (24)

[0128] Z Low =PointWiseConv(Z4)+Z (25)

[0129] (5) High-frequency component prediction module

[0130] A simple linear model is used to directly predict the high-frequency components in the power transformer data. Get the prediction results of the high-frequency component of the power transformer oil temperature data like Figure 4 As shown, this process can be expressed as shown in formula (26):

[0131] Res High =W Mlp r(t) (26)

[0132] (6) Prediction module

[0133] The feature Z extracted by the low-frequency component feature extraction module of the time series data Low After flattening, it is directly projected to the specified dimension and added to the prediction result obtained by the high-frequency component feature extraction module as the output result of the model. This process is shown in formula (27) to formula (29).

[0134] Z′ Low =Flatten(Z Low ) (27)

[0135] Res Low =W Mlp Z′ Low (28)

[0136]

[0137] Output the model results Scaling and moving are performed to supplement the unstable information deleted in the initial normalization module to generate the final input Y of the model. The specific process is shown in formula (30). Y is the final prediction result of the model for the oil temperature data of the power transformer, where E t [X] and Var[X] are defined the same as in formulas (2) and (3).

[0138]

[0139] In some embodiments, the prediction results are evaluated using two evaluation indicators, MSE and MAE, as shown in formulas (31) and (32).

[0140]

[0141]

[0142] Among them, y represents the actual value of the time series, Represents the predicted value, and n is the length of the predicted sequence.

[0143] The method provided in the above embodiment is compared with four methods in the prior art: K-nearest neighbor method (KNN), temporal convolutional neural network (TCN), gated recurrent unit neural network (GRU), and Transformer neural network. The two indicators of the method in the above embodiment are both the lowest, ranking first among the comparison methods, which confirms the effectiveness of the method in the above embodiment.

[0144] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0146] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0148] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made by those skilled in the art that fall within the spirit and principles of the present invention and do not require creative effort are intended to be within the scope of protection of the present invention.

Claims

1. A method for predicting oil temperature of a power transformer, characterized in that: The following steps are involved: Preprocessing the acquired power transformer oil temperature data; Decomposing the pre-processed power transformer oil temperature data to obtain high-frequency component and low-frequency component; The segmented attention mechanism is used instead of the dot product attention mechanism to improve the neural network model. The improved neural network model is used to extract the global and local features of the low-frequency component, and the extracted features are projected to the prediction length to obtain the first prediction result. Using a linear model to predict the high-frequency component of the data, a second prediction result is obtained; The first prediction result and the second prediction result are integrated to obtain the final prediction result.

2. A method for predicting oil temperature of a power transformer as claimed in claim 1, characterized in that: The preprocessing process includes normalization and denoising, wherein the acquired power transformer oil temperature data is normalized, and the normalized time series is converted from a time domain representation to a frequency domain using Fourier transform, and denoising is performed; Or further, remove the frequency band representing the noise in the frequency domain, retain the first K dimensions, and randomly sample the remaining dimensions, and select s components from the L Fourier components to save, where K, L, and s are all integers, and s <L; Or further, after the denoising data sequence is completed and filled, an inverse Fourier transform is performed to convert the frequency domain data back to the time domain wave.

3. A method for predicting oil temperature of a power transformer as claimed in claim 1, characterized in that: The process of decomposing the preprocessed power transformer oil temperature data includes decomposing the preprocessed power transformer oil temperature data using a variational mode decomposition method to obtain multiple modal components and residual terms, wherein the modal components are used as the low-frequency component part of the power transformer oil temperature data, and the residual terms are used as the high-frequency component part of the power transformer oil temperature data.

4. A method for predicting oil temperature of a power transformer as claimed in claim 1, characterized in that: The specific process of extracting the global and local features of the low-frequency component using the improved neural network model includes: The low-frequency component is segmented according to the set segment length and division step, and the modal component is divided into segment structures; The input of the segment structure is model embedded and position encoded; The low-frequency component of the power transformer oil temperature data is modeled through the Transformer model, and the features extracted from the low-frequency component of the power transformer oil temperature data are obtained.

5. A method for predicting oil temperature of a power transformer as claimed in claim 4, characterized in that: The modeling process of the low-frequency component of the power transformer oil temperature data using the Transformer model includes: Generate a linear projection query matrix, a key matrix and a value matrix about the input through linear projection; Split the input query matrix, key matrix, and value matrix according to the principle of multi-head attention; Calculate the attention distribution of each segment, concatenate the output results of all heads as the output of the multi-head attention, and then pass it through the normalization layer and the feedforward network layer with residual connection to obtain the intermediate modeling result; A convolutional structure is introduced in the hidden dimension to help capture local information within each segment, and the global and local features of the data are found through the global information network and the local information network respectively.

6. A method for predicting oil temperature of a power transformer as claimed in claim 1, characterized in that: The specific process of fusing the first prediction result and the second prediction result includes flattening the first prediction result and directly projecting it to the specified dimension, adding it to the second prediction result, and performing data denormalization on the added prediction result to obtain the final prediction result.

7. A power transformer oil temperature prediction system, characterized in that: include: a preprocessing module configured to preprocess the acquired power transformer oil temperature data; a decomposition module configured to decompose the preprocessed power transformer oil temperature data to obtain a high-frequency component part and a low-frequency component part; A low-frequency component feature extraction module is configured to use a segmented attention mechanism instead of a dot-product attention mechanism to improve the neural network model, extract global and local features of the low-frequency component using the improved neural network model, and project the extracted features onto a prediction length to obtain a first prediction result; a high-frequency component feature extraction module configured to predict the high-frequency component of the data using a linear model to obtain a second prediction result; The prediction module is configured to fuse the first prediction result and the second prediction result to obtain a final prediction result.

8. A power transformer oil temperature prediction system as claimed in claim 7, characterized in that: The preprocessing module includes a data normalization module and a denoising module. The data normalization module is configured to perform normalization processing on the acquired power transformer oil temperature data. The denoising module is configured to use Fourier transform to convert the normalized time series from a time domain representation to a frequency domain representation and perform denoising operation. Or further, the denoising module includes: A Fourier transform module is configured to transform the normalized time series from a representation in the time domain to a frequency domain; The noise removal module is configured to remove the frequency band representing the noise in the frequency domain, retain the first K dimensions, and randomly sample the remaining dimensions, selecting s components from the L Fourier components for storage, where K, L, and s are all integers, and s <L; A completion module, configured to complete and fill the output of the noise removal module; The inverse Fourier transform module is configured to perform inverse Fourier transform to convert the frequency domain data back into a time domain wave.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the steps of the method according to any one of claims 1 to 6.

10. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein the steps of the method according to any one of claims 1 to 6 are completed when the computer instructions are executed by the processor.

Images