Oil temperature prediction method for oil-immersed power transformer

Abstract

The application discloses an oil temperature prediction method of an oil-immersed power transformer, comprising the following steps: obtaining a historical operation data sequence of the oil-immersed power transformer; mapping the historical operation data sequence to a frequency domain by using frequency domain transformation, and decomposing the historical operation data sequence into at least two frequency band branch signals of different frequency components; respectively performing inverse transformation on each frequency band branch signal to reconstruct a frequency band time domain subsequence corresponding to the frequency band branch signal; inputting each frequency band time domain subsequence into a preset three-branch heterogeneous modeling model for feature extraction to obtain three feature outputs; adaptively weighting and fusing the obtained three feature outputs to obtain a fusion feature corresponding to the frequency band time domain subsequence; reconstructing and mixing all fusion features in a frequency channel dimension to obtain a unified time sequence representation, and outputting a transformer oil temperature prediction value at a future time through a prediction head. The method can realize high-precision and high-robustness long-time sequence prediction of the transformer oil temperature.

Description

Technical Field

[0001] This application relates to the field of transformer condition prediction, and more particularly to a method for predicting the oil temperature of an oil-immersed power transformer. Background Technology

[0002] Existing long-term prediction models for power transformer oil temperature mostly employ a uniform structure and fixed channel size for different frequency components. Long-term trends in low frequencies and subtle changes in high frequencies are treated in the same or two general branches, lacking specificity in capacity allocation and feature transformation. As a result, low-frequency components are easily characterized while high-frequency details are smoothed out, or too many parameters are acquired in high frequencies, making them sensitive to noise. Errors tend to accumulate more easily within long prediction windows. Furthermore, intra-band fusion methods are mostly static weighting or weak adaptive mechanisms, making it difficult to select a more suitable processing path based on the actual spectral distribution of the data. Stability and generalization ability are insufficient when crossing datasets and sampling frequencies. Summary of the Invention

[0003] This application provides a method for predicting the oil temperature of an oil-immersed power transformer, which enables high-precision, robust, and long-term time-series prediction of transformer oil temperature.

[0004] This application provides a method for predicting the oil temperature of an oil-immersed power transformer, the method comprising: Obtain the historical operating data sequence of the oil-immersed power transformer, which includes at least oil temperature information and transformer load characteristics; The historical operational data sequence is mapped to the frequency domain using frequency domain transformation and decomposed into frequency band branch signals with at least two different frequency components. Inverse transforms are performed on the signals of each frequency band branch to reconstruct the frequency band time domain subsequence of the corresponding frequency band branch signal; Each frequency band time domain subsequence is input into a preset three-branch heterogeneous modeling model for feature extraction, resulting in three feature outputs. The three-branch heterogeneous modeling model is configured to perform feature dimension mixing on the frequency band time domain subsequence through the first branch to obtain the first feature output, perform time dimension modeling on the frequency band time domain subsequence through the second branch to obtain the second feature output, and perform feature extraction on the frequency band time domain subsequence through the third branch to match the frequency distribution characteristics of the current frequency band time domain subsequence to obtain the third feature output. The obtained three feature outputs are adaptively weighted and fused to obtain the fused features of the corresponding frequency band time domain subsequence; All fusion features are reconstructed and mixed in the frequency channel dimension to obtain a unified time series representation, and the predicted transformer oil temperature value for future time moments is output through the prediction head.

[0005] In one embodiment, the first branch includes a gated linear mixer configured to process the input frequency band time domain subsequence in the order of dimension-up transformation, gated modulation, and dimension-down transformation to obtain a first feature output, or to process the input frequency band time domain subsequence in the order of gated modulation, dimension-up transformation, and dimension-down transformation to obtain a first feature output. The dimension upscaling transform expands the input sequence to a higher dimension space in terms of feature dimensions; the gated modulation uses a gated signal to modulate the received sequence position by position; the dimension reduction transform compresses the feature dimensions of the received sequence back to the original feature dimensions.

[0006] In one embodiment, processing the input frequency band time domain subsequence in the order of dimension up transformation, gated modulation, and dimension down transformation to obtain the first feature output includes: The input frequency band time domain subsequence is normalized. The feature vector of each time position in the normalized frequency band time domain subsequence is used as the processing object. The first linear transformation is performed along the feature dimension using the weight parameters shared at different time positions to expand the feature dimension to a preset multiple, thereby obtaining the first intermediate sequence. A gated signal is generated by a gated branch, and the first intermediate sequence is selectively modulated position by position using the gated signal to obtain the second intermediate sequence. The feature dimensions of the second intermediate sequence are compressed back to the original feature dimensions using a second linear transformation, and residual fusion is performed with the input frequency band time-domain subsequence to obtain the first feature output. In one embodiment, the gated branch is activated to generate a gated signal using any of the following gate functions: Sigmoid activation function, SiLU activation function, GELU activation function, gated residual activation function, multiply-accumulate mixed activation function, or temperature-modulated Softmax activation function.

[0007] In one embodiment, the second branch includes a channel-by-channel timing filter module, which is configured to: Receive frequency band time domain subsequences arranged according to "batch × channel × time"; For each feature channel in the frequency band time domain subsequence, a one-dimensional filtering operation along the time dimension is performed independently to capture the local dependencies between adjacent time steps. The one-dimensional filtering operation only uses left padding, and information mixing between different feature channels is prohibited during the filtering process. The output of the one-dimensional filtering operation is used as the second feature output to achieve time-dimensional modeling.

[0008] In one embodiment, the implementation structure of the channel-by-channel timing filtering module includes any of the following: A deep convolutional structure has a number of convolutional kernel groups equal to the number of channels of the input features, and each convolutional kernel operates only on the corresponding single feature channel; Grouped convolutional structures have the number of groups set to be equal to the number of channels of the input features, so that the convolution operation within each group only applies to the corresponding single feature channel; dynamic convolutional structures have convolutional kernel weights that are dynamically generated by the input features and adaptively adjust the filtering parameters for each channel or each time step. The depthwise separable convolutional structure includes at least a channel-wise convolutional layer, which performs the independent one-dimensional filtering operation; the depthwise separable convolutional structure also includes a pointwise mapping layer located after the channel-wise convolutional layer, which is used for output dimension mapping or subsequent feature integration.

[0009] In one embodiment, the third branch includes a frequency adapter, which is configured to: Based on the frequency range corresponding to the time domain subsequence of the frequency band, frequency adaptation operators with different time domain receptive field characteristics or different signal response characteristics are called to perform independent modeling. The configuration of the frequency adaptation operator is matched with the frequency range of the time-domain subsequence of the frequency band: For the frequency band time domain subsequence corresponding to the low frequency component, the first type of adaptation operator with long-range dependency capture capability is invoked to extract long-term trend features; For the frequency band time domain subsequence corresponding to the intermediate frequency component, the second type of adaptation operator with nonlinear morphological reconstruction capability is invoked to extract periodic or structural features. For the frequency band time domain subsequence corresponding to the high-frequency component, the third type of adaptation operator with transient change sensitivity is invoked to extract small fluctuations or edge features.

[0010] In one embodiment, the first type of adaptation operator uses any of the following algorithms to extract the long-term trend of low-frequency components: dilated convolution with large kernels, smooth spline fitting algorithm, low-pass filtering algorithm, trend term regression algorithm, and state-space long memory algorithm. The second type of adaptation operator uses any of the following algorithms to extract the periodic or structural features of the intermediate frequency components: Multilayer perceptron with nonlinear activation layer, gated residual MLP, channel shuffling MLP, frequency domain rearrangement module, subband attention module; The third type of adaptation operator uses any of the following algorithms to extract minute fluctuations or edge features of high-frequency components: The system is based on a second-order differential operator with repeated convolution, a high-pass filter module, an edge enhancement operator, a difference pyramid module, and a sparse spike detector.

[0011] In one embodiment, the obtained three feature outputs are adaptively weighted and fused to obtain the fused features of the corresponding frequency band time domain subsequence, including: Perform a global convergence operation on the first feature output, the second feature output, and the third feature in the time dimension to obtain a sample-level context description vector; The sample-level context description vector is input into the gated network, and through the mapping calculation of the gated network, three weight vectors corresponding to the three feature outputs are generated. The three weight vectors are broadcast-expanded along the time series dimension to match the dimension of the three feature outputs. Then, the expanded weight vectors are multiplied element-wise with the corresponding three feature outputs at each time position, and the results of the multiplications are summed to obtain the fused features of the time domain subsequence of the corresponding frequency band.

[0012] In one embodiment, the fusion features corresponding to time-domain subsequences of different frequency bands are aligned in the time dimension; Point-based mixing operations are applied to the frequency channel dimension formed by the aligned fused features. Through a set of learnable channel mixing weights, the multi-frequency features at each time position are linearly integrated into a unified time-series representation of a single path. The learnable channel mixing weights are applied consistently to all time positions within the same layer. The unified time series representation is input into the prediction head, which is configured as follows: Lightweight normalization is performed on the unified time-series representation to obtain normalized features; The transformer oil temperature prediction for future time periods is output through feature mapping using a fully connected network structure containing at least two layers. The fully connected network structure includes: The first linear layer is used to expand the dimension of the obtained normalized features to a high-dimensional hidden space; A nonlinear activation layer, connected after the first linear layer, performs nonlinear mapping on the output of the first linear layer to obtain activation features; The second linear layer projects the activation features onto the target output dimension, outputting the predicted transformer oil temperature for future timeframes. The aforementioned oil temperature prediction method for oil-immersed power transformers acquires historical operating data sequences including oil temperature information and transformer load characteristics. Through frequency domain transformation, the historical operating data sequences are decomposed into frequency band branch signals with different frequency components, and then reconstructed through inverse transformation to obtain the corresponding frequency band time domain subsequences. Each frequency band time domain subsequence is input into a preset three-branch heterogeneous modeling model. This model extracts three features through parallel feature extraction from the first, second, and third branches. Adaptive weighted fusion integrates the three features into fused features for each frequency band. All fused features of all frequency bands are then mixed and reconstructed along the frequency channel dimension to form a unified time series representation. The predicted transformer oil temperature value for future moments is output through a prediction head. This method achieves targeted processing of different frequency components and adaptive fusion of multi-dimensional features, thereby obtaining a more comprehensive understanding of oil temperature variation patterns and improving the accuracy of oil temperature prediction and adaptability to complex operating conditions. Attached Figure Description

[0013] Figure 1 This is a flowchart of an oil temperature prediction method for an oil-immersed power transformer in one embodiment; Figure 2 This is a flowchart of obtaining the first feature output in one embodiment; Figure 3 This is an architecture diagram of an oil temperature prediction method for an oil-immersed power transformer in one embodiment. Detailed Implementation

[0014] The present application will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present application. Any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the protection scope of the present application.

[0015] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0016] In one embodiment, such as Figure 1 As shown, a method for predicting the oil temperature of an oil-immersed power transformer is provided, which includes the following steps: Step 101: Obtain the historical operating data sequence of the oil-immersed power transformer. The historical operating data sequence shall include at least oil temperature information and transformer load characteristics. Specifically, historical operating data sequences can be collected by sensors to monitor the operation of oil-immersed transformers in actual power systems, recording load changes and oil temperature responses of oil-immersed transformers during long-term continuous operation, and reflecting the operating characteristics of oil-immersed transformers under different load levels and time periods.

[0017] Step 102: Map the historical running data sequence to the frequency domain using frequency domain transformation, and decompose it into frequency band branch signals with at least two different frequency components; Specifically, a Fast Fourier Transform is performed on the input historical operating data sequence (including oil temperature information OT and multiple transformer load characteristics) to map the historical operating data sequence from the time domain to the frequency domain, obtaining the corresponding spectral representation. Based on a preset number of channels, the spectrum is divided into frequency bands (for example, divided into three sub-bands from low to high frequency: low frequency, medium frequency, and high frequency), decomposing the spectrum into frequency band branches with at least two different frequency components.

[0018] It should be noted that the Fourier transform can also be equivalently replaced by discrete cosine, wavelet, short-time Fourier, S-transform, Hilbert transform, empirical mode / variable mode decomposition, learnable filter bank or hybrid transform; allowing complex domain or amplitude-phase separation representation.

[0019] Step 103: Perform inverse transform on each frequency band branch signal to reconstruct the frequency band time domain subsequence of the corresponding frequency band branch signal; Specifically, an inverse Fourier transform is performed on each frequency band branch signal to convert the frequency band branch signal from a frequency domain representation to a time domain representation. This allows for the reconstruction of a frequency band time domain subsequence that is aligned with the time length and sampling points of the historical running data sequence but carries different frequency components.

[0020] In this process, the input historical running data sequence is first decomposed into high / medium / low frequency bands by frequency domain transformation, and then inversely transformed back to the time domain to form a multi-path, complementary time series representation. Combined with timestamp embedding enhancement, it forms a stable frequency-time dual-domain representation framework.

[0021] Step 104: Input each frequency band time domain subsequence into the preset three-branch heterogeneous modeling model for feature extraction to obtain three feature outputs. The three-branch heterogeneous modeling model is configured to perform feature dimension mixing on the frequency band time domain subsequence through the first branch to obtain the first feature output, perform time dimension modeling on the frequency band time domain subsequence through the second branch to obtain the second feature output, and perform feature extraction on the frequency band time domain subsequence through the third branch to match the frequency distribution characteristics of the current frequency band time domain subsequence to obtain the third feature output. Specifically, a Fast Fourier Transform (FFT) is performed on the input historical data sequence to map it from the time domain to the frequency domain, obtaining the corresponding spectral representation. The spectrum is then divided into low-frequency, mid-frequency, and high-frequency band signals, from low to high frequency. Inverse Fourier Transforms are then performed on the low-frequency, mid-frequency, and high-frequency band signals respectively, yielding the reconstructed low-frequency, mid-frequency, and high-frequency band time-domain subsequences.

[0022] The reconstructed low-frequency, mid-frequency, and high-frequency band time-domain subsequences are input into a pre-defined three-branch heterogeneous modeling model for feature extraction. This three-branch heterogeneous modeling model can initiate three branches in parallel for each frequency band time-domain subsequence. The first branch may include a gated linear mixer, which can perform feature dimension mixing on the time-domain subsequences of each frequency band by first linearly increasing the feature dimension, then selectively passing through the position by gating activation, and finally linearly reducing the feature dimension and adding it to the input residual to obtain the first feature output.

[0023] The second branch processes the input frequency band temporal subsequence through a channel-wise one-dimensional convolution module. This module performs temporal filtering on each feature channel independently, strictly maintaining the information independence between the various feature channels. Temporal feature capture is performed on each feature channel, modeling the time step, and thus outputting a second-path feature focused on temporal variations.

[0024] The third branch may include a frequency adapter, which is configured to call a dedicated frequency channel matching the frequency range of each frequency band time-domain subsequence for feature extraction. For example, dilated convolutions with larger kernels and larger receptive fields are used for low-frequency frequency band time-domain subsequences; multilayer perceptrons with nonlinear activation layers are used for temporal feature rearrangement for mid-frequency frequency band time-domain subsequences; and second-order differencing is used for high-frequency frequency band time-domain subsequences to capture small abrupt changes. The result of the frequency adapter extracting features from each frequency band time-domain subsequence is the third feature output.

[0025] Step 105: Adaptively weighted fuse the obtained three feature outputs to obtain the fused features of the corresponding frequency band time domain subsequence; Specifically, the contextual information of the three features is globally aggregated to form a description vector that can characterize the overall characteristics of the current sample. This description vector is input into a gating network, which then dynamically calculates and generates a set of non-negative fusion weights corresponding to the three features. Finally, the fusion weights are applied to the corresponding three features, and fusion is completed through weighted summation to obtain the fusion features of the corresponding frequency band time domain subsequence.

[0026] Step 106: Reconstruct and mix all the fused features in the frequency channel dimension to obtain a unified time series representation, and output the predicted value of transformer oil temperature at future time through the prediction head.

[0027] Specifically, the fused features of each frequency band are ensured to be aligned in the time dimension, i.e., having the same time step and resolution. Point-based mixing operations are applied to the multi-band features at each time position in the frequency channel dimension; that is, the multi-band features are linearly integrated into a single-path feature sequence through a set of learnable channel mixing weights. These mixing weights remain consistent across all time positions within the same reconstruction layer, thus ensuring the stability and consistency of the reconstruction mixing process. Finally, a unified time-series representation fusing all frequency band information is output.

[0028] The unified time series representation is input into the prediction head, which maps the input unified time series representation to the target output dimension, namely the predicted value of transformer oil temperature at future time, through a multi-layer structure of "linear transformation-nonlinear activation-linear projection", thereby completing the prediction from historical operating data sequence to future oil temperature.

[0029] In this embodiment, the method obtains frequency band time-domain subsequences with the same frequency components by performing frequency domain decomposition and inverse transform on historical operating data sequences. Each frequency band time-domain subsequence is then subjected to feature dimension mixing, time dimension modeling, and feature extraction matching frequency distribution characteristics through a three-branch heterogeneous modeling model. These subsequences are then subjected to adaptive weighted fusion and frequency channel dimension reconstruction and fusion, and finally, the predicted oil temperature value is output through the prediction head. This method achieves targeted processing of different frequency components and adaptive fusion of multi-dimensional features, thereby obtaining a more comprehensive understanding of oil temperature change patterns, improving the accuracy of oil temperature prediction, and enhancing adaptability to complex operating conditions.

[0030] In one embodiment, the first branch includes a gated linear mixer configured to: process the input frequency band time domain subsequence in the order of dimension-up transformation, gated modulation, and dimension-down transformation to obtain a first feature output; or process the input frequency band time domain subsequence in the order of gated modulation, dimension-up transformation, and dimension-down transformation to obtain a first feature output. The dimension upscaling transform expands the input sequence to a higher dimension space in terms of feature dimensions; the gated modulation uses a gated signal to modulate the received sequence position by position; the dimension reduction transform compresses the feature dimensions of the received sequence back to the original feature dimensions.

[0031] Specifically, the first branch includes a gated linear mixer configured to process the input frequency band time-domain subsequence to obtain a first feature output. The gated linear mixer includes two configurations: The first method involves dimensionality upscaling, gated modulation, and dimensionality reduction. Dimensional upscaling involves lightly normalizing the input frequency band time-domain subsequence and then applying a linear transformation along the feature dimension at each time position, temporarily expanding the feature dimension to a preset multiple (e.g., twice) of the original feature dimension, thus extending the input subsequence into a higher-dimensional space. Gated modulation generates a gate signal through parallel gated branches, and based on the gate signal, selectively passes through and modulates the aforementioned high-dimensional frequency band time-domain subsequence position by position. Dimensional reduction involves compressing the high-dimensional feature representation back to the original feature dimension through another linear transformation, and finally adding it to the initial input frequency band time-domain subsequence through a residual path to obtain the first feature output.

[0032] The second method involves gated modulation, dimensionality increase, and dimensionality decrease. A gated linear mixer processes the input frequency band time-domain subsequence in the order of gated modulation, dimensionality increase, and dimensionality decrease to obtain the first feature output. Specifically, it first uses a gate function to modulate the input frequency band time-domain subsequence position-by-position, and then performs linear dimensionality increase and decrease transformations.

[0033] In addition, the gating activation function used in gating modulation is GELU. Other activation functions can be equivalently replaced by Sigmoid activation function, SiLU activation function, gated residual activation function, multiply-accumulate hybrid activation function, or temperature-modulated Softmax activation function, etc.

[0034] In one embodiment, such as Figure 2 As shown, the input frequency band time domain subsequence is processed in the order of dimension-up transformation, gated modulation, and dimension-down transformation to obtain the first feature output, including the following steps: Step 201: Normalize the input frequency band time domain subsequence, take the feature vector of each time position in the normalized frequency band time domain subsequence as the processing object, and perform the first linear transformation along the feature dimension using the weight parameters shared at different time positions to expand the feature dimension to a preset multiple to obtain the first intermediate sequence. Step 202: Generate a gated signal through a gated branch, and use the gated signal to selectively modulate the first intermediate sequence position by position to obtain the second intermediate sequence; Step 203: Use the second linear transformation to compress the feature dimension of the second intermediate sequence back to the original feature dimension, and perform residual fusion with the input frequency band time domain subsequence to obtain the first feature output.

[0035] Specifically, a lightweight normalization process is performed on the input frequency band time domain subsequence, and the feature vector at each time position in the normalized frequency band time domain subsequence is used as the processing object. A first linear transformation is performed along the feature dimension using weight parameters shared across different time positions. This first linear transformation is a dimension-up transformation, which expands the feature dimension of the frequency band time domain subsequence to a preset multiple (e.g., 2x or 4x) of the original feature dimension, thus obtaining the first intermediate sequence.

[0036] A gated signal is generated through a gated branch set in parallel with the aforementioned linear transformation branch; this gated branch can use GELU (Gaussian Error Linear Unit) as the activation function. The first intermediate sequence is then selectively modulated position-by-position using this gated signal, i.e., each position element of the high-dimensional intermediate representation is dynamically weighted to control the throughput, thereby obtaining the modulated second intermediate sequence.

[0037] The second linear transformation is used to compress the feature dimension of the second intermediate sequence back to the feature dimension of the original input sequence, completing the dimensionality reduction transformation. The result of the dimensionality reduction transformation is added to and fused with the initial input frequency band time domain subsequence (through a residual path), finally outputting the first feature output.

[0038] In one embodiment, the second branch includes a channel-by-channel timing filter module, which is configured to: Receive frequency band time domain subsequences arranged according to "batch × channel × time"; For each feature channel in the frequency band time domain subsequence, a one-dimensional filtering operation along the time dimension is performed independently to capture the local dependencies between adjacent time steps. The one-dimensional filtering operation only uses left padding, and information mixing between different feature channels is prohibited during the filtering process. The output of the one-dimensional filtering operation is used as the second feature output to achieve time-dimensional modeling.

[0039] Specifically, the channel-by-channel temporal filtering module receives the input frequency band time-domain subsequences in a "batch × channel × time" format. This module employs one-dimensional convolution to perform operations along the time dimension. By setting the `groups` parameter of the one-dimensional convolution to be equal to the number of feature channels in the input frequency band time-domain subsequence, it ensures that each feature channel in the frequency band time-domain subsequence can undergo independent one-dimensional filtering, and prohibits information mixing between different feature channels during the one-dimensional filtering process. Furthermore, the output of the one-dimensional filtering operation is used as the second feature output to achieve temporal dimension modeling.

[0040] In addition, the channel-wise temporal filtering module can capture the local dependencies of adjacent time steps by configuring explicit convolution kernel size, stride and padding, and introduce dilated convolution to systematically expand the receptive field to model longer-distance temporal dimensions.

[0041] It should be noted that, to meet the causality requirements of time series prediction, one-dimensional filtering operations only use left-side padding. That is, when performing convolutional filtering on the sequence along the time dimension, a specific number of padding values ​​(usually 0) are added only at the beginning (left) of the time axis of each input sequence, without adding any padding at the end (right) of the sequence. The purpose is to ensure that when performing feature calculations or output predictions at each time step t, the input information relied upon comes entirely from the current time t and previous historical times, ensuring that the prediction behavior is consistent with the time series logic.

[0042] In one embodiment, the implementation structure of the channel-by-channel timing filtering module includes any of the following: A deep convolutional structure has a number of convolutional kernel groups equal to the number of channels of the input features, and each convolutional kernel operates only on the corresponding single feature channel; Grouped convolutional structures have the number of groups set to be equal to the number of channels of the input features, so that the convolution operation within each group only applies to the corresponding single feature channel; dynamic convolutional structures have convolutional kernel weights that are dynamically generated by the input features and adaptively adjust the filtering parameters for each channel or each time step. A depthwise separable convolutional structure includes at least a channel-wise convolutional layer that performs the independent one-dimensional filtering operation; the depthwise separable convolutional structure may also include a pointwise mapping layer located after the channel-wise convolutional layer, the pointwise mapping layer being used for output dimension mapping or subsequent feature integration.

[0043] Specifically, deep convolutional structures can set the number of groups in a one-dimensional convolution to be equal to the number of channels in the input features, thus making the number of convolutional kernel groups equal to the number of channels in the input features. Under this setting, each convolutional kernel operates only on its corresponding single feature channel, achieving independent processing of feature channels and thereby capturing the local dependencies of each feature channel in the temporal dimension.

[0044] Grouped convolutional structures can set the number of groups to be equal to the number of channels in the input features. In this case, each group corresponds to only one feature channel, and the convolution operation is performed only within that single feature channel along the time dimension. Its effect is consistent with channel-wise depthwise convolution, satisfying the design goal of preventing information mixing between different feature channels. Dynamic convolutional structures do not have fixed kernel weights; they are dynamically generated from the input features. This allows the filtering parameters to be adaptively adjusted for each sample, each feature channel, and even each time step, thereby enhancing adaptability to changes in input features.

[0045] Depthwise separable convolutional structures include at least a channel-wise convolutional layer. The channel-wise convolutional layer functions essentially the same as the depthwise convolutional structure described above, performing independent one-dimensional filtering operations along the time dimension. A pointwise mapping layer (1x1 convolution) can be placed after the channel-wise convolutional layer as a post-processing layer for output dimension mapping or subsequent feature integration; this pointwise mapping layer is not part of the aforementioned independent one-dimensional filtering operation itself.

[0046] In one embodiment, the third branch includes a frequency adapter, which is configured to: Based on the frequency range corresponding to the time domain subsequence of the frequency band, frequency adaptation operators with different time domain receptive field characteristics or different signal response characteristics are called to perform independent modeling. The configuration of the frequency adaptation operator is matched with the frequency range of the time-domain subsequence of the frequency band: For the frequency band time domain subsequence corresponding to the low frequency component, the first type of adaptation operator with long-range dependency capture capability is invoked to extract long-term trend features; For the frequency band time domain subsequence corresponding to the intermediate frequency component, the second type of adaptation operator with nonlinear morphological reconstruction capability is invoked to extract periodic or structural features. For the frequency band time domain subsequence corresponding to the high-frequency component, the third type of adaptation operator with transient change sensitivity is invoked to extract small fluctuations or edge features.

[0047] Specifically, the frequency adapter adaptively calls frequency adaptation operators with different time-domain receptive field characteristics or different signal response characteristics for independent modeling based on the frequency range corresponding to the time-domain subsequence of the frequency band. The configuration of each operator is strictly matched with the frequency range of the time-domain subsequence of the frequency band to achieve targeted feature extraction. For the time-domain subsequence of the frequency band corresponding to the low-frequency component, the first-class adaptation operator with long-range dependency capture capability is invoked. The first-class adaptation operator uses dilated convolution with a large convolution kernel to effectively capture slow changes and long-term trend features in the signal by utilizing the large receptive field.

[0048] For the time-domain subsequence of the frequency band corresponding to the mid-frequency component, a second type of adaptation operator with nonlinear morphological reconstruction capability is invoked. The second type of adaptation operator uses a multilayer perceptron with a nonlinear activation function (such as GELU) to rearrange and transform the time-dimensional features in order to extract their periodic or structural features.

[0049] For the time-domain subsequence of the frequency band corresponding to the high-frequency components, a third type of adaptation operator with transient change sensitivity is invoked. The third type of adaptation operator uses second-order difference operation or repeated convolution model to enhance the sensitivity to rapidly changing signals, so as to effectively extract the edge features of small fluctuations or changes.

[0050] In this embodiment, by calling the first, second and third types of adaptation operators on the time domain subsequences of low, medium and high frequency bands respectively, long-term trends, periodic patterns and transient changes can be accurately captured, reducing feature confusion and long-term prediction error accumulation caused by the one-size-fits-all treatment of different frequency components.

[0051] In one embodiment, the first type of adaptation operator is used to extract the long-term trend of low-frequency components, and the main implementation algorithm is dilated convolution with a large kernel. Alternatively, the implementation algorithm can be equivalently replaced by a smooth spline fitting algorithm, a low-pass filtering algorithm, a trend term regression algorithm, or a state-space long memory algorithm.

[0052] The second type of adaptation operator is used to extract periodic or structural features of the intermediate frequency components. The main implementation algorithm is a multilayer perceptron with a nonlinear activation layer. The implementation algorithm can also be replaced by a gated residual MLP, a channel shuffling MLP, a frequency domain rearrangement module, or a subband attention module.

[0053] The third type of adaptation operator is used to extract minute fluctuations or edge features of high-frequency components. The main implementation algorithm is a second-order differential operator based on repeated convolution. In addition, the implementation algorithm can also be replaced by a high-pass filter module, an edge enhancement operator, a difference pyramid module, or a sparse spike detector.

[0054] In one embodiment, the obtained three feature outputs are adaptively weighted and fused to obtain the fused features of the corresponding frequency band time domain subsequence, including: Perform a global convergence operation on the first feature output, the second feature output, and the third feature in the time dimension to obtain a sample-level context description vector; The sample-level context description vector is input into the gated network, and through the mapping calculation of the gated network, three weight vectors corresponding to the three feature outputs are generated. The three weight vectors are broadcast-expanded along the time series dimension to match the dimension of the three feature outputs. Then, the expanded weight vectors are multiplied element-wise with the corresponding three feature outputs at each time position, and the results of the multiplications are summed to obtain the fused features of the time domain subsequence of the corresponding frequency band.

[0055] Specifically, the first, second, and third feature outputs from the gated linear mixer, channel-wise one-dimensional convolution, and frequency adapter are subjected to a global average convergence operation in the time dimension to obtain a sample-level context description vector representing global context information.

[0056] Furthermore, the sample-level context description vector is input into a gating network (e.g., a network consisting of fully connected layers and a softmax function). Through the mapping calculation of this gating network, three non-negative weight vectors that sum to 1 are generated, corresponding to the three feature outputs. These three weight vectors are then broadcast-expanded along the time series dimension to match the dimensions of the three weight vectors with the dimensions of the three feature outputs.

[0057] Furthermore, the expanded weight vector is multiplied element-wise with the corresponding three feature outputs at each time position, and the results of the multiplications are summed to obtain the fusion features of the corresponding frequency band time domain subsequence.

[0058] It should be noted that the Softmax function can also be equivalently replaced by the temperature-controlled Softmax function, the entmax function, the normalized ReLU / Softsign function, the sigmoid scaling function, and the simple linear weighted and summed normalized function; it allows the generation of sparse or hard selection functions (Top-k / sampling).

[0059] In one embodiment, the fusion features corresponding to time-domain subsequences of different frequency bands are aligned in the time dimension; Point-based mixing operations are applied to the frequency channel dimension formed by the aligned fused features. Through a set of learnable channel mixing weights, the multi-frequency features at each time position are linearly integrated into a unified time-series representation of a single path. The learnable channel mixing weights are applied consistently to all time positions within the same layer.

[0060] Specifically, the fusion features corresponding to the time-domain subsequences from the low-frequency, mid-frequency, and high-frequency bands are aligned in the time dimension to ensure that different fusion features have the same number of time steps and length. If the temporal resolution is inconsistent due to differences in step size or receptive field in the upstream processing branches, alignment can be performed before mixing through a learnable resampling operation to ensure that the synthesized time-series trajectory is free of artifacts and phase shift.

[0061] In the frequency channel dimension formed by the aligned and fused features, a point-based mixing operation is applied. This involves linearly integrating multiple frequency features at each time point into a unified time-series representation of a single path using a set of learnable channel mixing weights. These learnable channel mixing weights are consistently applied across all time points within the same layer, ensuring the stability and reproducibility of the synthesis process. After this frequency band mixing is completed, a unified time-series representation incorporating information from the entire frequency band is obtained.

[0062] It should be noted that the aforementioned fusion weights can be dynamically generated adaptively based on the input fusion feature samples, enabling the adaptive adjustment of the importance of each branch's output features for different input conditions. Furthermore, the use of the Softmax function to normalize the weights ensures the stability of the fusion process and the interpretability of the results.

[0063] In this embodiment, the fusion features of multiple frequency bands are integrated into a unified timing representation through alignment and blending. This preserves the complete information of the entire frequency band, avoids information loss or conflicts that may occur during manual fusion, and ensures the stability and efficiency of the reconstruction process.

[0064] In one embodiment, a unified time-series representation is input into the prediction head, and the prediction head is configured as follows: Lightweight normalization is performed on the unified time-series representation to obtain normalized features; The transformer oil temperature prediction for future time periods is output through feature mapping using a fully connected network structure containing at least two layers. The fully connected network structure includes: The first linear layer is used to expand the dimension of the obtained normalized features to a high-dimensional hidden space; A nonlinear activation layer, connected after the first linear layer, performs nonlinear mapping on the output of the first linear layer to obtain activation features; The second linear layer is used to project the activation features onto the target output dimension, outputting the predicted transformer oil temperature at future time points.

[0065] Specifically, a lightweight normalization process is performed on the input unified temporal representation to obtain normalized features; the normalized features are then input into a fully connected network structure containing at least two layers for feature mapping, the fully connected network structure may include a first linear layer, a nonlinear activation layer and a second linear layer.

[0066] The first linear layer is used to expand the dimension of the normalized features to a high-dimensional hidden space; the nonlinear activation layer is connected after the first linear layer and performs nonlinear mapping processing on the output of the first linear layer, for example, by using activation functions such as GELU to obtain activation features; the second linear layer is used to project the activation features onto the target output dimension, that is, to directly output the predicted value of transformer oil temperature at a specified future time.

[0067] As a preferred embodiment, the framework diagram of the oil temperature prediction method for oil-immersed power transformers is as follows: Figure 3As shown. In the frequency decomposition stage, historical operating data sequences containing oil temperature (OT) information and multi-dimensional transformer load characteristics are obtained as input sequences. A Fast Fourier Transform (FFT) is performed on the historical operating data sequences to decompose them into three frequency band branches: high frequency, medium frequency, and low frequency. An Inverse Fourier Transform is then performed on the frequency band signals for each frequency to reconstruct the corresponding frequency band time-domain subsequences. These subsequences are then embedded and enhanced with contextual information such as timestamps to form three enhanced frequency band time-domain subsequences.

[0068] Furthermore, in the multi-channel representation learning stage, the aforementioned high-frequency, mid-frequency, and low-frequency band sub-sequences are respectively input into a preset three-branch heterogeneous modeling model for feature extraction.

[0069] First branch: The gated linear mixer performs "linear-gated-linear" mixing along the sequence dimension; The second branch: Channel-wise convolution performs temporal modeling along the time dimension to capture local dependencies; The third branch: The frequency adaptation unit adaptively calls second-order difference / repeated convolution, nonlinear MLP, or large receptive field dilated convolution operators respectively, depending on whether the current processed frequency band subsequence is high, medium, or low frequency, in order to match its frequency characteristics.

[0070] The outputs of the three branches are weighted and fused to generate the fused features for that frequency band. During the frequency mixing stage, the three fused features from the high, medium, and low frequency channels are aligned in the time dimension, and point-wise mixing operations are applied in the frequency channel dimension. These features are then linearly integrated into a unified time-series representation using a set of learnable weights. This unified time-series representation is input into a perceptron (i.e., a multilayer perceptron prediction head), which outputs the predicted transformer oil temperature for future time moments.

[0071] Based on the above-mentioned oil temperature prediction method for oil-immersed power transformers, this application takes the historical operating data sequence including oil temperature information and transformer load characteristics as input, takes the multi-dimensional feature sequence of several consecutive time steps as input window, and takes the oil temperature sequence of several time steps after the window as prediction target to achieve multi-step forward oil temperature prediction.

[0072] To verify the effectiveness of the proposed method, this application compares the proposed method (TimeGate) with related methods (including itransformer, PatchTST, Dlinear, FEDformer, and Autoformer) and tests it on publicly available datasets. The ElectricityTransformer Temperature (ETT) dataset used in this application's embodiments originates from industrial data on transformer operation monitoring in actual power systems, recording load changes and oil temperature response during long-term continuous operation. This dataset realistically reflects the operating characteristics of transformers under different load levels and time periods, and can be used to construct and validate various models for transformer oil temperature prediction and operating status assessment.

[0073] The ETT dataset comprises four subsets: ETTh1, ETTh2, ETTm1, and ETTm2. ETTh1 and ETTh2 are hourly sampled subsets, while ETTm1 and ETTm2 are 15-minute sampled subsets. Each subset contains raw fields in a uniform format: a timestamp (date) and seven numerical variables (HUFL, HULL, MUFL, MULL, LUFL, LULL, and OT). HUFL, HULL, MUFL, MULL, LUFL, and LULL are predefined transformer load characteristics in the ETT dataset, corresponding to two types of load measurement variables under high, medium, and low load levels, respectively. OT represents transformer oil temperature, which is the primary prediction target of this application. Therefore, at any given time step, the data sample's feature dimension consists of seven numerical variables plus a timestamp field.

[0074] At the spatial and equipment level, ETTh1 and ETTm1 originate from monitoring data of the same transformer, differing only in their sampling intervals of 1 hour and 15 minutes, respectively; ETTh2 and ETTm2 originate from another transformer, also providing hourly and 15-minute observation sequences, respectively. By simultaneously introducing hourly and 15-minute subsets, this application can model and compare the dynamic relationship between transformer temperature and load at different time resolutions, thereby verifying the applicability and robustness of the method under different sampling densities and operating conditions.

[0075] In terms of modeling, this application uses the aforementioned multidimensional load characteristics and historical oil temperature sequence as model input, and uses the future oil temperature (OT) as model output for prediction. Specifically, a multidimensional feature sequence of several consecutive time steps can be used as an input window, and the oil temperature sequence of several time steps after this window can be used as the prediction target to achieve multi-step look-ahead oil temperature prediction.

[0076] By training and validating on four subsets, ETTh1, ETTh2, ETTm1 and ETTm2, the generalization ability and prediction accuracy of the method on different transformer objects and different time scales can be systematically evaluated.

[0077]

[0078] Based on this table, it can be concluded that for most datasets, the oil temperature prediction method (Timegate) proposed in this application achieves optimal performance compared to five other state-of-the-art time series prediction models (iTransformer, PatchTST, DLinear, FEDformer, and Autoformer) across various prediction lengths. Even for the ETTh2 dataset, the performance of the oil temperature prediction method (Timegate) proposed in this application is only slightly lower than that of iTransfermer, but for other time series prediction models, the oil temperature prediction method (Timegate) proposed in this application achieves the best performance.

[0079] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0080] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A method for predicting the oil temperature of an oil-immersed power transformer, characterized in that, The oil temperature prediction method includes the following steps: Obtain the historical operating data sequence of the oil-immersed power transformer, wherein the historical operating data sequence includes at least oil temperature information and transformer load characteristics; The historical running data sequence is mapped to the frequency domain using frequency domain transformation and decomposed into frequency band branch signals with at least two different frequency components; Each of the frequency band branch signals is subjected to an inverse transform to reconstruct the corresponding frequency band time domain subsequence of the frequency band branch signal; Each frequency band time-domain subsequence is input into a preset three-branch heterogeneous modeling model for feature extraction to obtain three feature outputs. The three-branch heterogeneous modeling model is configured to perform feature dimension mixing on the frequency band time-domain subsequence through the first branch to obtain the first feature output, perform time dimension modeling on the frequency band time-domain subsequence through the second branch to obtain the second feature output, and perform feature extraction on the frequency band time-domain subsequence through the third branch to match the frequency distribution characteristics of the current frequency band time-domain subsequence to obtain the third feature output. The obtained three feature outputs are adaptively weighted and fused to obtain the fused features of the corresponding frequency band time domain subsequence; All the fused features are reconstructed and mixed in the frequency channel dimension to obtain a unified time series representation, and the predicted transformer oil temperature value for future time is output through the prediction head. The first branch includes a gated linear mixer, which is configured to: process the input frequency band time domain subsequence in the order of dimension-up transformation, gated modulation, and dimension-down transformation to obtain the first feature output; or process the input frequency band time domain subsequence in the order of gated modulation, dimension-up transformation, and dimension-down transformation to obtain the first feature output. The up-dimensional transformation expands the input sequence to a higher-dimensional space in terms of feature dimensions; the gated modulation uses a gated signal to modulate the received sequence position by position; the down-dimensional transformation compresses the feature dimensions of the received sequence back to the original feature dimensions.

2. The method for predicting the oil temperature of an oil-immersed power transformer according to claim 1, characterized in that, The first feature output is obtained by processing the input frequency band time domain subsequence in the order of dimension-up transformation, gated modulation, and dimension-down transformation. The input frequency band time domain subsequence is normalized. The feature vector of each time position in the normalized frequency band time domain subsequence is used as the processing object. The first linear transformation is performed along the feature dimension using the weight parameters shared at different time positions to expand the feature dimension to a preset multiple, thereby obtaining the first intermediate sequence. A gated signal is generated by a gated branch, and the first intermediate sequence is selectively modulated position by position using the gated signal to obtain a second intermediate sequence; The feature dimension of the second intermediate sequence is compressed back to the original feature dimension using a second linear transformation, and residual fusion is performed with the input frequency band time domain subsequence to obtain the first feature output.

3. The method for predicting the oil temperature of an oil-immersed power transformer according to claim 1, characterized in that, The gated branch is activated to generate the gated signal using any of the following gated functions: Sigmoid activation function, SiLU activation function, GELU activation function, gated residual activation function, multiply-accumulate mixed activation function, or temperature-modulated Softmax activation function.

4. The method for predicting the oil temperature of an oil-immersed power transformer according to claim 1, characterized in that, The second branch includes a channel-by-channel timing filter module, which is configured as follows: The frequency band time domain subsequence is received according to the arrangement of "batch × channel × time"; For each feature channel in the time-domain subsequence of the frequency band, a one-dimensional filtering operation along the time dimension is performed independently to capture the local dependencies between adjacent time steps. The one-dimensional filtering operation only uses left padding and prohibits the mixing of information between different feature channels during the filtering process. The output of the one-dimensional filtering operation is used as the second feature output to achieve the time-dimensional modeling.

5. The method for predicting the oil temperature of an oil-immersed power transformer according to claim 4, characterized in that, The implementation structure of the channel-by-channel time-series filtering module includes any of the following: A deep convolutional structure has a number of convolutional kernel groups equal to the number of channels of the input features, and each convolutional kernel operates only on the corresponding single feature channel; Grouped convolutional structures have the number of groups set to be equal to the number of channels of the input features, so that the convolution operation within each group only applies to the corresponding single feature channel; dynamic convolutional structures have convolutional kernel weights that are dynamically generated by the input features and adaptively adjust the filtering parameters for each channel or each time step. The depthwise separable convolutional structure includes at least a channel-wise convolutional layer, which performs the independent one-dimensional filtering operation; the depthwise separable convolutional structure also includes a pointwise mapping layer located after the channel-wise convolutional layer, which is used for output dimension mapping or subsequent feature integration.

6. The method for predicting the oil temperature of an oil-immersed power transformer according to claim 1, characterized in that, The third branch includes a frequency adapter, which is configured to: Based on the frequency range corresponding to the time domain subsequence of the frequency band, frequency adaptation operators with different time domain receptive field characteristics or different signal response characteristics are called to perform independent modeling. The configuration of the frequency adaptation operator is matched with the frequency range of the frequency band time domain subsequence. For the frequency band time domain subsequence corresponding to the low frequency component, the first type of adaptation operator with long-range dependency capture capability is invoked to extract long-term trend features; For the frequency band time domain subsequence corresponding to the intermediate frequency component, the second type of adaptation operator with nonlinear morphological reconstruction capability is invoked to extract periodic or structural features. For the frequency band time domain subsequence corresponding to the high-frequency component, the third type of adaptation operator with transient change sensitivity is invoked to extract small fluctuations or edge features.

7. The method for predicting the oil temperature of an oil-immersed power transformer according to claim 6, characterized in that, The first type of adaptation operator uses any of the following algorithms to extract the long-term trend of low-frequency components: dilated convolution with large convolution kernel, smooth spline fitting algorithm, low-pass filtering algorithm, trend term regression algorithm, and state space long memory algorithm. The second type of adaptation operator uses any of the following algorithms to extract the periodic or structural features of the intermediate frequency components: Multilayer perceptron with nonlinear activation layer, gated residual MLP, channel shuffling MLP, frequency domain rearrangement module, subband attention module; The third type of adaptation operator uses any of the following algorithms to extract minute fluctuations or edge features of high-frequency components: The system is based on a second-order differential operator with repeated convolution, a high-pass filter module, an edge enhancement operator, a difference pyramid module, and a sparse spike detector.

8. The method for predicting the oil temperature of an oil-immersed power transformer according to claim 1, characterized in that, The step of adaptively weighting and fusing the obtained three feature outputs to obtain the fused features of the corresponding frequency band time domain sub-sequence includes: Perform a global convergence operation on the first feature output, the second feature output, and the third feature in the time dimension to obtain a sample-level context description vector; The sample-level context description vector is input into a gating network, and through the mapping calculation of the gating network, three weight vectors corresponding to the three feature outputs are generated. The three weight vectors are broadcast-expanded along the time series dimension to match the dimension of the three feature outputs. Then, the expanded weight vectors are multiplied element-wise with the corresponding three feature outputs at each time position, and the results of the multiplications are summed to obtain the fusion features of the time domain subsequence of the corresponding frequency band.

9. The method for predicting the oil temperature of an oil-immersed power transformer according to claim 1, characterized in that, Align the fusion features corresponding to time-domain subsequences of different frequency bands in the time dimension; Point-based mixing operations are applied on the frequency channel dimension formed by the aligned fused features. Through a set of learnable channel mixing weights, the multi-frequency features at each time position are linearly integrated into a unified time-series representation of a single path. The learnable channel mixing weights are applied consistently to all time positions within the same layer. The unified time series representation is input into the prediction head, and the prediction head is configured as follows: Lightweight normalization is performed on the unified time-series representation to obtain normalized features; The transformer oil temperature prediction for future times is output through feature mapping using a fully connected network structure comprising at least two layers. The fully connected network structure includes: The first linear layer is used to expand the dimension of the obtained normalized features to a high-dimensional hidden space; A nonlinear activation layer is connected after the first linear layer, and the output of the first linear layer is nonlinearly mapped to obtain activation features. The second linear layer is used to project the activation features onto the target output dimension to output the predicted transformer oil temperature at future times.

Images