Power Transformer Temperature Prediction Method and System Based on Multi-Feature Extraction and Fusion

Abstract

The present invention discloses a power transformer temperature prediction method and system based on multi-feature extraction and fusion, which perform spatio-temporal decoupling standardization and causal emergence clustering on each temperature variable in the data to obtain a clustering result; extract common patterns from each class of the clustering result to obtain a multi-variable time series with information complemented and downsample it to obtain multiple subsequences; respectively perform feature extraction on the multi-variable time series with information complemented and the subsequences through a dynamic time-frequency domain MLP mechanism to obtain overall features and local features, and perform fusion to obtain time series features; obtain external environmental covariates and input them into a self-attention mechanism network to obtain the self-features of the covariates, and perform feature fusion on the self-features of the covariates and the time series features through a covariate-driven causal convolution fusion mechanism to obtain fusion features; input the fusion features into a dynamic attention feature weighted prediction model for prediction to obtain a prediction result. It significantly improves the accuracy and reliability of power transformer temperature prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment status monitoring, and particularly to a power transformer temperature prediction method and system based on multi-feature extraction and fusion. Background Art

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

[0003] With the deepening of the construction of the smart grid, the power system's demand for the status monitoring and fault warning of key equipment is becoming increasingly urgent. As the core equipment of the power grid, the operating temperature of a power transformer directly reflects the health status and load capacity of the equipment. Accurately predicting the temperature change of the transformer not only helps to optimize the equipment operation and maintenance strategy and extend the service life, but also can effectively prevent power outages caused by overheating faults, which is of great significance for ensuring power supply reliability and promoting the digital transformation of energy.

[0004] The temperature of a power transformer is dynamically coupled and affected by multi-dimensional features such as winding temperature, oil temperature, radiator temperature, and ambient temperature, and has significant nonlinearity and time series correlation. Existing data-driven temperature prediction methods still have obvious limitations in feature mining and modeling, mainly manifested in the following problems:

[0005] (1) When dealing with multiple temperature variables, some existing methods usually simply superimpose all other temperature variables without considering the correlation between variables, which easily leads to the introduction of noise by the interaction between irrelevant variables, thus affecting the prediction effect.

[0006] (2) Previous methods have not fully characterized the dynamic interaction mechanism between multi-source features. They mainly capture the features of the internal temperature variables of the power transformer, but do not introduce the dynamic interaction between external environmental covariates, resulting in the prediction effect being limited by a bottleneck.

[0007] (3) Common time series modeling methods (such as long short-term memory neural network (LSTM) and temporal convolutional network (TCN)) are insufficient in terms of time series dependence and complex feature extraction ability, and are difficult to effectively capture and extract complex feature relationships. Summary of the Invention

[0008] To overcome the above-mentioned deficiencies of the prior art, the present invention provides a power transformer temperature prediction method and system based on multi-feature extraction and fusion, which significantly improves the accuracy and reliability of power transformer temperature prediction.

[0009] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0010] In a first aspect, the present invention provides a power transformer temperature prediction method based on multi-feature extraction and fusion, including:

[0011] Obtain multivariate time series data of a power transformer composed of temperature variables;

[0012] Perform spatio-temporal decoupling normalization on the multivariate time series data, and perform causal emergence clustering on each temperature variable in the normalized multivariate time series data to obtain a clustering result;

[0013] Extract common patterns from each class of the clustering result, splice and fuse the common patterns into each class sequence to obtain a multivariate time series with complemented information;

[0014] Downsample the multivariate time series with complemented information to obtain multiple subsequences; extract overall features from the multivariate time series with complemented information through a dynamic time-frequency domain MLP mechanism, extract local features from the subsequences through a dynamic time-frequency domain MLP mechanism, and fuse the overall features and local features to obtain time series features;

[0015] Obtain external environmental covariates and input them into a self-attention mechanism network to obtain covariate self-features, and fuse the covariate self-features and time series features through a covariate-driven causal convolution fusion mechanism to obtain fused features;

[0016] Input the fused features into a dynamic attention feature weighted prediction model for prediction to obtain a prediction result.

[0017] In a further technical solution, the specific clustering result is:

[0018] Perform spatio-temporal decoupling normalization on the multivariate time series data;

[0019] Calculate the mutual information between each temperature variable in the normalized multivariate time series data, and convert the mutual information into a similarity matrix;

[0020] Obtain a symmetric Laplacian matrix based on the similarity matrix, perform eigenvalue decomposition on the symmetric Laplacian matrix to obtain a feature matrix;

[0021] Use a causal emergence clustering algorithm to cluster the feature matrix to obtain a clustering result.

[0022] In a further technical solution, using a causal emergence clustering algorithm to cluster the feature matrix specifically is:

[0023] According to the feature matrix, learn the causal DAG structure between variables by optimizing the objective function to construct a causal graph;

[0024] Input the feature matrix and the causal graph into a graph neural network for feature extraction to obtain causal features;

[0025] Calculate the cosine similarity of the causal features, and calculate the causal similarity based on the cosine similarity;

[0026] Use the causal similarity as the clustering criterion, and incorporate causal constraints in the hierarchical clustering process to obtain the clustering result.

[0027] For a further technical solution, the multivariate time series with information complemented is specifically obtained as follows:

[0028] Adaptive convolutional pooling is used to obtain the significant features of all temperature variables within each category, and random pooling is used to obtain the overall features of all temperature variables within each category;

[0029] Fuse the significant features and the overall features of each category respectively, and splice the fused results to each temperature variable within the corresponding category;

[0030] Use a long short-term memory network to perform feature fusion on each spliced temperature variable to obtain the multivariate time series with information complemented.

[0031] For a further technical solution, both the global features and the local features are extracted through a dynamic time-frequency domain MLP mechanism, specifically:

[0032]

[0033]

[0034]

[0035] Among them, represents the input sequence, represents the result of frequency domain MLP modeling, represents the continuous wavelet transform, represents the sequence transformed back to the time domain, represents the inverse continuous wavelet transform, represents the sparse MLP, represents the output features after passing through the dynamic time-frequency domain MLP mechanism.

[0036] For a further technical solution, the fused features are specifically obtained as follows: Obtain the dynamic convolutional kernel at each time step based on the features of the covariate itself, align the features of the covariate itself and the temporal features according to the time step, and perform causal convolutional fusion according to the dynamic convolutional kernel at each time step.

[0037] For a further technical solution, in the dynamic attention feature weighted prediction model, a single-layer lightweight attention network is embedded in linear regression to achieve feature-level dynamic adjustment.

[0038] In a second aspect, the present invention provides a power transformer temperature prediction system based on multi-feature extraction and fusion, including:

[0039] A data acquisition module, which is configured to: acquire multivariate time series data composed of temperature variables of a power transformer;

[0040] A similarity clustering module, which is configured to: perform spatio-temporal decoupling normalization on the multivariate time series data, and perform causal emergence clustering on each temperature variable in the normalized multivariate time series data to obtain a clustering result;

[0041] An intra-class common pattern extraction module, which is configured to: extract common patterns from each class of the clustering result, splice and fuse the common patterns into each class sequence to obtain a multivariate time series with complemented information;

[0042] A time series feature modeling module, which is configured to: downsample the multivariate time series with complemented information to obtain multiple subsequences; extract overall features from the multivariate time series with complemented information through a dynamic time-frequency domain MLP mechanism, extract local features from the subsequences through a dynamic time-frequency domain MLP mechanism, and fuse the overall features and local features to obtain time series features;

[0043] A covariate interaction module, which is configured to: acquire external environmental covariates and input them into a self-attention mechanism network to obtain covariate self-features, and fuse the covariate self-features and time series features through a covariate-driven causal convolution fusion mechanism to obtain fusion features;

[0044] A multi-step prediction module, which is configured to: input the fusion features into a dynamic attention feature weighted prediction model for prediction to obtain a prediction result.

[0045] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the steps in the power transformer temperature prediction method based on multi-feature extraction and fusion as described in the first aspect.

[0046] In a fourth aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the steps in the power transformer temperature prediction method based on multi-feature extraction and fusion as described in the first aspect.

[0047] The above one or more technical solutions have the following beneficial effects:

[0048] The present invention uses a causal emergence clustering algorithm based on mutual information to perform correlation clustering on each temperature variable, effectively reducing the influence of irrelevant information; extracts the common patterns of variables through a combination of random pooling and adaptive convolutional pooling to ensure the completion of missing or ambiguous information and enhance the stability of the prediction model; uses a covariate-driven causal convolution fusion mechanism and a self-attention mechanism to achieve deep feature fusion between external environmental covariates and temperature variables, significantly improving the accuracy of multi-source feature interaction; finally, performs multi-step prediction on the temperature through a dynamic attention feature weighted prediction model, significantly improving the accuracy and reliability of power transformer temperature prediction.

[0049] When standardizing the multivariate time series data composed of temperature variables of the power transformer, the present invention considers the dimensional differences between variables and introduces a spatio-temporal decoupling standardization technique to separately decouple time and space standardization, solving the problem of multivariate coupling interference.

[0050] The present invention designs a causal emergence clustering algorithm combining causal discovery and hierarchical clustering to perform correlation clustering on temperature variables, identifies the causal dependencies between variables in the data, and uses this causal information to guide the clustering process, so that the formation of clusters is not only based on data similarity but also reflects the causal mechanism between variables.

[0051] The present invention proposes a method of feature extraction in the form of a dynamic time-frequency domain MLP mechanism (TFMLP). Compared with the conventional MLP, it can not only process frequency domain features but also reduces the complexity of the traditional MLP through sparsification, realizing the full modeling of time-frequency domain features.

[0052] When predicting the temperature of the power transformer, the present invention embeds a single-layer lightweight attention network into linear regression to achieve feature-level dynamic adjustment, enhancing the prediction ability for complex scenarios containing multiple temperature variables. Description of the Drawings

[0053] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0054] Figure 1 is the flowchart of data correlation clustering in the temperature prediction method of the embodiment of the present invention;

[0055] Figure 2 is the flowchart of common pattern extraction in the temperature prediction method of the embodiment of the present invention;

[0056] Figure 3 is the flowchart of time series feature modeling in the temperature prediction method of the embodiment of the present invention;

[0057] Figure 4It is the flow chart of dynamic time-frequency domain MLP modeling in the temperature prediction method of the embodiment of the present invention;

[0058] Figure 5 It is the flow chart of covariate feature interaction in the temperature prediction method of the embodiment of the present invention;

[0059] Figure 6 It is the module diagram of the temperature prediction system of the embodiment of the present invention. Detailed implementation manners

[0060] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0061] It should be noted that the terms used herein are only for describing specific implementation manners 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 forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0062] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0063] Embodiment 1

[0064] This embodiment discloses a power transformer temperature prediction method based on multi-feature extraction and fusion, and the method includes the following steps:

[0065] S1: Obtain multi-variable time series data composed of temperature variables of the power transformer;

[0066] In this embodiment, the temperature of the power transformer is dynamically coupled and affected by multi-dimensional features such as winding temperature, oil temperature, radiator temperature, and ambient temperature. Each temperature variable index of the power transformer, such as oil temperature, winding temperature, radiator temperature, etc., can be regarded as a variable. Therefore, the data of each index constitutes multi-variable time series data, which is defined as , where represents the number of temperature variables of the power transformer, represents the length of the time series data.

[0067] S2: Perform spatio-temporal decoupling normalization on the multi-variable time series data, and perform causal emergence clustering on each temperature variable in the normalized multi-variable time series data to obtain a clustering result;

[0068] In this embodiment, in order to avoid the excessive difference in the numerical range between different variables from affecting the clustering effect, the multivariate time series data is standardized. In data standardization, the current standardization only considers the difference characteristics in the time dimension of the time series data, but does not consider the dimensional difference between variables. Therefore, the present invention introduces a spatio-temporal decoupling standardization technique to perform spatio-temporal decoupling standardization on the multivariate time series data, decoupling time and space standardization respectively to solve the problem of multivariate coupling interference. Specifically, first, standardize in the time dimension to eliminate the temporal drift of a single variable: The expression is:

[0069] (1)

[0070] Among them, represents the result of standardization in the time dimension, represents the data in the mean value in the time dimension, represents the data in the variance in the time dimension.

[0071] Then standardize in the space dimension to eliminate the dimensional difference between variables:

[0072] (2)

[0073] (3)

[0074] (4)

[0075] (5)

[0076] (6)

[0077] Among them, represents the number of power transformer temperature variables, represents the th power transformer temperature variable, represents the th variable at the th time step, represents the data each variable at time when the mean value, at the th time step, the result of spatial normalization, represents the result of coupling time normalization and spatial normalization, represents the data each variable at time when the variance, represents the scaling factor, Denotes the offset. The scaling factor and offset are added after normalization as an affine transformation to enhance the expressive power of the model.

[0078] In this embodiment, as Figure 1 shown, the mutual information clustering technique is used to cluster according to the correlation between variables, clustering highly correlated variables together. This method first calculates the mutual information between variables, replacing the traditional Pearson correlation coefficient, to capture the non-linear dependence relationship between variables. Define the mutual information , where denotes the number of power transformer temperature variables, and the mutual information calculation formula is as follows:

[0079] (7)

[0080] Among them, denotes the th row of the mutual information matrix, denotes the th column of the mutual information matrix, denotes the th variable data, denotes the th variable data, denotes the value of the th variable at the th time step, denotes the value of the th variable at the th time step, is the joint / marginal probability distribution of this variable.

[0081] To enhance the similarity discrimination, the exponential function Gaussian kernel is used to transform the mutual information value to obtain the similarity matrix , expressed as:

[0082] (8)

[0083] Among them, is a positive constant used to control the attenuation speed of the similarity. A larger will make the similarity more sensitive to the change of the mutual information value.

[0084] Then, according to the similarity matrix construct the degree matrix , is a diagonal matrix, and the elements are the degrees of each node; then obtain the Laplacian matrix according to the degree matrix, and obtain the symmetric Laplacian matrix according to the Laplacian matrix; then perform eigenvalue decomposition on the symmetric Laplacian matrix to obtain the top The eigenvectors corresponding to the smallest eigenvalues form an eigenmatrix . The above calculation expression is as follows:

[0085] (9)

[0086] (10)

[0087] (11)

[0088] (12)

[0089] Among them, represents the eigenvalue, represents the number of temperature variables of the power transformer. Each row of the eigenmatrix is regarded as the low-dimensional representation of a variable.

[0090] For the clustering method of the eigenmatrix, the present invention designs a causal emergence clustering algorithm that combines causal discovery and hierarchical clustering. The causal emergence clustering algorithm is used to cluster the eigenmatrix to obtain a clustering result. Specifically: according to the eigenmatrix, the causal DAG structure between variables is learned by optimizing the objective function to construct a causal graph; the eigenmatrix and the causal graph are input into a graph neural network for feature extraction to obtain causal features; the cosine similarity of the causal features is calculated, and the causal similarity is calculated based on the cosine similarity; the causal similarity is used as the clustering criterion, and causal constraints are incorporated in the hierarchical clustering process to obtain a clustering result.

[0091] In traditional clustering methods, the causal dependence relationship between variables is not considered, which will affect the clustering effect. Therefore, the present invention identifies the causal dependence relationship between variables in the data and uses this causal information to guide the clustering process, so that the formation of clusters is not only based on data similarity but also reflects the causal mechanism between variables.

[0092] Specifically, according to the eigenmatrix , the causal DAG structure between variables is learned by optimizing the objective function:

[0093] (13)

[0094] Among them, represents the adjacency matrix of the causal graph, representing the causal weight between variables, that is, the causal strength of variable on variable ; represents the reconstruction loss based on the graph structure; represents the constraint term of the directed acyclic graph, , Indicates dynamic weight balance, sparsity, and acyclicity, avoiding overfitting; Indicates the Frobenius norm, and its calculation method is the square root of the sum of the squares of the absolute values of all elements in the matrix; Indicates the sparsity regularization term, which controls the complexity (sparsity) of the causal graph.

[0095] Next, based on the causal graph (i.e., the adjacency matrix ), causal features are extracted to enhance the interpretability of clustering:

[0096] (14)

[0097] where, Indicates the causal feature, Indicates the graph neural network, Indicates the activation function, Indicates the degree matrix, Indicates the feature matrix, Indicates the learnable parameter, Indicates the adjacency matrix with self-connections added.

[0098] Finally, the causal similarity is defined as the cluster merging criterion:

[0099] (15)

[0100] where, Indicates the causal similarity; Indicates the balance weight between the data similarity and the causal path contribution; Indicates the causal feature of the cosine similarity, used to measure the similarity between variable and variable in the causal feature space; Indicates the set of all causal paths from variable to variable , Indicates the causal effect of path .

[0101] During the hierarchical clustering process, causal constraints are incorporated to form a macroscopic cluster structure, obtaining the final clustering result. This introduction of causal information guidance in hierarchical clustering can break through the limitations of traditional data-driven methods and further improve the accuracy of clustering.

[0102] Through the above technical solution, the mutual information clustering is used to perform correlation clustering on each temperature variable of the power transformer, clustering highly correlated variables together and avoiding the influence of irrelevant noise.

[0103] S3: Extract the common patterns from each class of the clustering results, splice and fuse the common patterns into the sequences of each class to obtain a multivariate time series with complemented information;

[0104] In this embodiment, as Figure 2 shown, after clustering, the common patterns include significant features and overall features. The method of random pooling combined with adaptive convolutional pooling is used to aggregate the features of each variable within the class. Specifically, random pooling considers globally and extracts the overall features within the class, while adaptive convolutional pooling performs convolution on the variables within the class to extract the significant features within the class. This method first uses adaptive convolutional pooling to obtain the most significant features of all variables within each class. At the same time, random pooling is used to obtain the overall features of all variables within each class. The calculation formulas are as follows:

[0105] (16)

[0106] (17)

[0107] Among them, represents the overall features within the class, represents the significant features within the class, represents the th class of all variables, represents random pooling, represents adaptive convolution. Both random pooling and adaptive convolutional pooling have the characteristics of self - adaptation. Compared with traditional average pooling and max pooling, they have stronger generalization ability.

[0108] Subsequently, the two types of features, namely the significant features and overall features of each class, are fused. After fusion, they are spliced onto the sequence of each variable within the class, and then feature fusion is performed through a shared LSTM to obtain a multivariate time series with complemented information , 。Feature fusion refers to the fusion of each variable within the class with the fused significant features and overall features. That is to say, after each variable within the class is spliced with the corresponding fused significant features and overall features, each variable is fused separately to incorporate the significant features and overall features of the class into the variable.

[0109] The modeling of LSTM (Long Short - Term Memory Network) is as follows:

[0110] (18)

[0111] (19)

[0112] (20)

[0113] (21)

[0114] (22)

[0115] (23)

[0116] Among them, represents the forget gate, represents the sigmoid activation function, , , , respectively represent the weight matrices of the forget gate, input gate, candidate memory unit, and output gate, , , respectively represent the bias terms of the forget gate, input gate, and candidate memory unit, represents the hidden state at the current time, represents the input at the current time, represents the input gate, represents the candidate memory unit, represents the memory unit at the current time, represents the output gate, represents the multiplication operation.

[0117] The purpose of doing this is to achieve the effective interaction of common patterns among highly correlated variables within a class, while avoiding the introduction of irrelevant noise due to the interaction between low-correlation variables.

[0118] Through the above technical solution, a method combining random pooling and adaptive convolutional pooling is used to extract common patterns from each class, and these common patterns are supplemented into the original sequence of variables within the class through the method of splicing and fusion, so as to complete the complement of missing or fuzzy information. The information missing or fuzzy here refers to the variables in the clustering. In S2, the relevant temperature variables are clustered together. A certain variable may lack the information contained in other variables. Fusing the common pattern within the class into this variable can achieve the supplement of the information of other variables, that is, information complementation.

[0119] S4: Downsample the multivariate time series with information complementation to obtain multiple subsequences; extract the overall features from the multivariate time series with information complementation through the dynamic time-frequency domain MLP mechanism, extract the local features from the subsequences through the dynamic time-frequency domain MLP mechanism, and fuse the overall features and local features to obtain the time series features;

[0120] In this embodiment, as Figure 3As shown, the temperature data of the power transformer (original time series), as time series data, has an inherent low-rank property. Therefore, the present invention performs information completion on the multivariate time series using uniform step downsampling to obtain multiple subsequences with a time pattern extremely similar to that of the sequence before downsampling , where represents the number of intervals. Feature extraction is performed on different subsequences respectively, which can more deeply mine the information in the multivariate time series for information completion , thereby enhancing the ability to extract time features. In this regard, the present invention proposes a method of a dynamic time-frequency domain MLP mechanism (TFMLP). As shown, this mechanism first uses wavelet transform to transfer the sequence to the frequency domain, and performs MLP (multi-layer perceptron) processing on the amplitude and phase in the frequency domain, and then returns to the time domain through inverse wavelet transform; further, sparse MLP is used for processing in the time domain to capture time domain features, and the sparse MLP means that only the Top-K neurons are activated and the rest are set to zero. The specific steps of the dynamic time-frequency domain MLP mechanism are as follows: Figure 4 As shown, this mechanism first uses wavelet transform to transfer the sequence to the frequency domain, and performs MLP (multi-layer perceptron) processing on the amplitude and phase in the frequency domain, and then returns to the time domain through inverse wavelet transform; further, sparse MLP is used for processing in the time domain to capture time domain features, and the sparse MLP means that only the Top-K neurons are activated and the rest are set to zero. The specific steps of the dynamic time-frequency domain MLP mechanism are as follows:

[0121] (24)

[0122] (25)

[0123] (26)

[0124] Among them, represents the input sequence, represents the modeling result of the frequency domain MLP, represents the continuous wavelet transform, represents the sequence returned to the time domain, represents the continuous inverse wavelet transform, represents the sparse MLP, represents the output features after passing through the dynamic time-frequency domain MLP mechanism. The dynamic time-frequency domain MLP mechanism (TFMLP) can not only process frequency domain features compared with the conventional MLP, but also reduces the complexity of the traditional MLP through sparsification, realizing the full modeling of time-frequency domain features.

[0125] The multivariate time series with information completion is modeled by one layer of TFMLP to obtain the overall features , and the subsequences are modeled by one layer of TFMLP to obtain the local features . After is spliced according to the position during downsampling, and then added and fused with to obtain the time series features . The expression is:

[0126] (27)

[0127] (28)

[0128] Among them, represents the overall feature, represents the th local feature.

[0129] S5: Obtain external environmental covariates and input them into the self-attention mechanism network to obtain the covariate self-features. Through the covariate-driven causal convolution fusion mechanism, fuse the covariate self-features with the temporal features to obtain the fused features; apply the covariate self-features as the "convolution weight" to the causal convolution, so that the convolution operation has dynamic adaptability in time series.

[0130] Specifically, the covariate self-features generate convolution kernel parameters through a network at each time step, so that the "receptive field" of the convolution operation of the main variable at each time step can be adaptively adjusted according to the state of the covariates. This method not only retains the structural information of the temporal features, but also makes the influence of the covariates on the main variable more flexible and dynamic.

[0131] In this embodiment, as Figure 5 shown, define the external environmental covariate data as , apply the self-attention mechanism modeling to the external environmental covariate data to obtain the covariate self-features . The external environmental covariates include, but are not limited to, external factors such as environmental temperature, air humidity, wind speed, sunshine intensity, and transformer load current that dynamically affect the heat dissipation efficiency. These covariates are collected in real time through sensors and aligned with the temperature variable time series to jointly form a multi-source feature input. Considering the external covariates will further improve the prediction accuracy. The self-attention mechanism modeling process is expressed as:

[0132] (29)

[0133] (30)

[0134] (31)

[0135] (32)

[0136] Among them, represents the external environmental covariates, represents the query vector, , , represent learnable weight parameters, Represents the key vector, Represented as a value vector, Represents the dot product of the query vector and the key vector, Represents the dimension of the key vector.

[0137] Then, using the designed covariate-driven causal convolution fusion mechanism, the covariate's own features are fused into the temporal features. First, the covariate's own feature data at the current time step is input into a fully connected neural network to obtain a dynamic convolution kernel:

[0138] (33)

[0139] Among them, Represents the Convolution kernel weight at the Represents the non-linear activation function, Represents the At the Represents the Learnable weight parameter at the Represents the Learnable bias term at the

[0140] Align the temporal feature And the covariate's own feature According to the time step to obtain the aligned feature . In each time step, Both contain the temporal feature and the covariate's own feature. Then, using the obtained dynamic convolution kernel, causal convolution fusion is performed on the temporal feature and the covariate's own feature contained in each time step, that is, the covariate information is fused into the temporal feature using causal convolution, expressed as:

[0141] (34)

[0142] Among them, Represents the final fused feature, Represents causal convolution.

[0143] Through the above technical solution, a uniform step size downsampling feature extraction method is used to model the temporal features of each temperature variable of the power transformer, and the external environment covariates and temperature variables are feature-fused through a covariate-driven causal convolution fusion mechanism and a self-attention mechanism.

[0144] S6: Input the fused feature into a dynamic attention feature weighted prediction model for prediction to obtain a prediction result.

[0145] In this embodiment, compared with single-step prediction and traditional linear regression multi-step prediction, the present invention introduces a dynamic attention feature weighted prediction model (which can also be called a multi-step prediction model), embeds a single-layer lightweight attention network into linear regression, realizes feature-level dynamic adjustment, and improves the prediction ability for complex scenarios containing multiple temperature variables.

[0146] Specifically, first, according to the feature fusion result at the current time step , calculate the hidden layer features:

[0147] (35)

[0148] Among them, represents the hidden layer features, represents the non-linear activation function, represents the weight matrix of the hidden layer, represents the bias term of the hidden layer, and respectively represent the mean and variance of

[0149] Next, generate attention-enhanced weights based on the hidden layer features:

[0150] (36)

[0151] Among them, represents the attention weight at time , represents the weight matrix of the query part, represents the weight matrix of the key part, represents the dimension of the input features, used to scale the attention scores; represents the weight matrix of the output part of the attention mechanism, represents the bias term of the output part of the attention mechanism, represents the non-linear activation function for calculating the attention weights.

[0152] Finally, perform weighted output based on the attention-enhanced weights:

[0153] (37)

[0154] Among them, is the prediction result of the power transformer temperature, represents the weight matrix of the weighted output, represents the bias term of the weighted output.

[0155] To better demonstrate the effectiveness of the method of the present invention, several common regression problem evaluation metrics are used to evaluate the results. These metrics include the mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE), and continuous ranked probability score (CRPS). MSE, MAE, and RMSE are point estimation evaluation metrics for single prediction results, while CRPS is used to evaluate the confidence probability estimation of multiple prediction results. The smaller the values of these metrics, the closer the prediction results are to the true values, and the better the prediction effect.

[0156] (38)

[0157] (39)

[0158] (40)

[0159] (41)

[0160] Among them, represents the true temperature value of the power transformer at time , represents the predicted temperature value of the method of the present invention at time , is the prediction length, represents the probability distribution of the temperature values of multiple point predictions.

[0161] The method of the present invention is compared with three advanced methods: long short-term memory neural network (LSTM), temporal convolutional neural network (TCN), and multi-head attention mechanism network (Multi-Head Attention). The method of the present invention has the best effect among all methods, and the error values of the evaluation metrics are the lowest, verifying the effectiveness of the proposed method.

[0162] Example Two

[0163] As Figure 6 shown, this embodiment discloses a power transformer temperature prediction system based on multi-feature extraction and fusion, including:

[0164] A data acquisition module, which is configured to: acquire multi-variable time series data of the power transformer composed of temperature variables;

[0165] A similarity clustering module, which is configured to: perform spatio-temporal decoupling normalization on the multi-variable time series data, and perform causal emergence clustering on each temperature variable in the normalized multi-variable time series data to obtain a clustering result;

[0166] The in-class common pattern extraction module is configured to: extract common patterns from each class of the clustering results, splice and fuse the common patterns into various class sequences to obtain a multi-variable time series with complemented information;

[0167] The time series feature modeling module is configured to: downsample the multi-variable time series with complemented information to obtain multiple subsequences; extract overall features from the multi-variable time series with complemented information through a dynamic time-frequency domain MLP mechanism, extract local features from the subsequences through a dynamic time-frequency domain MLP mechanism, and fuse the overall features and local features to obtain time series features;

[0168] The covariate interaction module is configured to: obtain external environmental covariates and input them into a self-attention mechanism network to obtain covariate self-features, and fuse the covariate self-features and time series features through a covariate-driven causal convolution fusion mechanism to obtain fused features;

[0169] The multi-step prediction module is configured to: input the fused features into a dynamic attention feature weighted prediction model for prediction to obtain prediction results.

[0170] Embodiment III

[0171] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the method in Embodiment I are implemented.

[0172] Embodiment IV

[0173] The purpose of this embodiment is to provide a computer-readable storage medium. A computer-readable storage medium stores a computer program, and when the program is executed by a processor, the steps of the method in Embodiment I are executed.

[0174] The steps involved in the devices in the above Embodiments III and IV correspond to those in Method Embodiment I. For specific implementation manners, reference may be made to the relevant description part of Embodiment I. The term "computer-readable storage medium" should be understood to include a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0175] Those skilled in the art should understand that the various modules or steps of the present invention described above can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by the computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0176] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0177] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A power transformer temperature prediction method based on multi-feature extraction and fusion, characterized in that: include: Obtain multivariate time series data consisting of temperature variables for power transformers; Performing spatiotemporal decoupling standardization on the multivariate time series data, and performing causal emergence clustering on each temperature variable in the standardized multivariate time series data to obtain a clustering result; Extracting common patterns from each class of the clustering results, splicing and fusing the common patterns into various classes of sequences, and obtaining a multivariate time series with complete information; Downsampling the information-completed multivariate time series to obtain multiple subsequences; extracting features of the information-completed multivariate time series through a dynamic time-frequency domain MLP mechanism to obtain overall features, extracting features of the subsequences through a dynamic time-frequency domain MLP mechanism to obtain local features, and fusing the overall features and local features to obtain time series features; Obtain external environmental covariates and input them into the self-attention mechanism network to obtain the covariate's own features, and fuse the covariate's own features with the time series features through the covariate-driven causal convolution fusion mechanism to obtain fused features; The fused features are input into the dynamic attention feature weighted prediction model for prediction to obtain a prediction result.

2. The power transformer temperature prediction method based on multi-feature extraction and fusion according to claim 1 is characterized in that: The clustering results are as follows: Standardize multivariate time series data by temporal and spatial decoupling; Calculating the mutual information between each temperature variable in the standardized multivariate time series data, and converting the mutual information into a similarity matrix; A symmetric Plaus matrix is ​​obtained based on the similarity matrix, and eigenvalue decomposition is performed on the symmetric Plaus matrix to obtain a characteristic matrix; The causal emergence clustering algorithm is used to cluster the feature matrix to obtain a clustering result.

3. The power transformer temperature prediction method based on multi-feature extraction and fusion as claimed in claim 2 is characterized in that: The causal emergence clustering algorithm is used to cluster the feature matrix as follows: According to the feature matrix, the causal DAG structure between variables is learned by optimizing the objective function to construct a causal graph; Inputting the feature matrix and the causal graph into a graph neural network for feature extraction to obtain causal features; Calculating the cosine similarity of the causal features, and obtaining the causal similarity based on the cosine similarity; The causal similarity is used as a clustering criterion, and causal constraints are incorporated into the hierarchical clustering process to obtain a clustering result.

4. The power transformer temperature prediction method based on multi-feature extraction and fusion according to claim 1 is characterized in that: The multivariate time series with completed information is as follows: Adaptive convolutional pooling is used to obtain the significant features of all temperature variables within each category, and random pooling is used to obtain the overall features of all temperature variables within each category; The significant features and overall features of each class are fused separately, and then spliced ​​to each temperature variable in the corresponding class; The long short-term memory network is used to fuse the features of each temperature variable after splicing to obtain a multivariate time series with complete information.

5. The power transformer temperature prediction method based on multi-feature extraction and fusion according to claim 1 is characterized in that: The overall features and local features are extracted through the dynamic time-frequency domain MLP mechanism, specifically: in, represents the input sequence, represents the frequency domain MLP modeling results, represents continuous wavelet transform, represents the sequence converted back to the time domain, represents the inverse continuous wavelet transform, represents the sparse MLP, Represents the output features after the dynamic time-frequency domain MLP mechanism.

6. The power transformer temperature prediction method based on multi-feature extraction and fusion according to claim 1 is characterized in that: The fusion features are obtained specifically as follows: based on the characteristics of the covariate itself, the dynamic convolution kernel of each time step is obtained, the characteristics of the covariate itself and the time series characteristics are aligned according to the time step, and causal convolution fusion is performed according to the dynamic convolution kernel at each time step.

7. The power transformer temperature prediction method based on multi-feature extraction and fusion according to claim 1 is characterized in that: The dynamic attention feature weighted prediction model embeds a single-layer lightweight attention network into linear regression to achieve dynamic adjustment at the feature level.

8. The power transformer temperature prediction system based on multi-feature extraction and fusion is characterized by: include: A data acquisition module is configured to: acquire multivariate time series data consisting of temperature variables of the power transformer; A similarity clustering module is configured to: perform spatiotemporal decoupling standardization on the multivariate time series data, perform causal emergence clustering on each temperature variable in the standardized multivariate time series data, and obtain a clustering result; An intra-class common pattern extraction module is configured to: extract common patterns from each class of the clustering results, and splice and fuse the common patterns into various classes of sequences to obtain a multivariate time series with complete information; A time series feature modeling module is configured to: downsample the information-completing multivariate time series to obtain multiple subsequences; extract features of the information-completing multivariate time series through a dynamic time-frequency domain MLP mechanism to obtain overall features, extract features of the subsequences through a dynamic time-frequency domain MLP mechanism to obtain local features, and fuse the overall features and local features to obtain time series features; A covariate interaction module is configured to: obtain external environment covariates and input them into the self-attention mechanism network to obtain covariate features, and fuse the covariate features with time series features through a covariate-driven causal convolution fusion mechanism to obtain fused features; The multi-step prediction module is configured to: input the fusion feature into the dynamic attention feature weighted prediction model for prediction to obtain a prediction result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the power transformer temperature prediction method based on multi-feature extraction and fusion as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the power transformer temperature prediction method based on multi-feature extraction and fusion as described in any one of claims 1 to 7 are implemented.

Images