Hybrid model-based power transformer periodic time sequence prediction method and multivariable periodic time sequence prediction method

Abstract

The invention discloses a power transformer periodic time sequence prediction method and a multivariable periodic time sequence prediction method based on a hybrid model, and the method comprises the steps: collecting the operation data of a transformer, including oil temperature and load variables, and carrying out the standardization, missing value processing and periodic time feature embedding; constructing a multi-branch hybrid model, fusing a bidirectional state space encoder, a time convolution network, a local window attention mechanism and a periodical enhancement network, and respectively capturing long-term periodical dependence, multi-scale local periodical features, a short-term periodical mode and periodical feature enhancement representation; the multi-branch features are adaptively integrated through a global context weighted fusion mechanism, efficient nonlinear transformation is performed on the fusion features by adopting a multi-order Chebyshev polynomial projection module, and a prediction sequence of oil temperature and load variables is generated.

Description

Technical Field

[0001] The present invention relates to the field of power system equipment state prediction, and in particular to a power transformer periodic time series prediction method and a multivariable periodic time series prediction method based on a hybrid model. Background Art

[0002] my country's energy structure is rapidly transitioning toward a clean, low-carbon future. This goal requires the power system to significantly increase the proportion of renewable energy (such as wind and solar) connected to the grid while optimizing energy efficiency to reduce carbon emissions and ensure energy security. However, renewable energy generation is intermittent and volatile, resulting in highly dynamic and cyclical power flow and load variations in the power system. This places higher demands on the operational stability and energy dispatch efficiency of power transformers.

[0003] As the core equipment of the power system, the oil temperature and load variables of power transformers (such as high-voltage side load, medium-voltage side load, and low-voltage side load) usually show obvious daily, weekly and seasonal periodic characteristics. These periodic patterns directly affect power dispatching, load balancing and equipment maintenance efficiency. In the context of dual carbon, accurate prediction of transformer oil temperature and load variables will not only help optimize new energy consumption and power dispatching, but also reduce energy consumption and carbon emissions by planning equipment maintenance in advance. However, traditional transformer operating parameter prediction methods mainly rely on simple statistical models or manual experience, which makes it difficult to effectively capture complex nonlinear relationships and multi-scale periodic characteristics, resulting in insufficient prediction accuracy and difficulty in meeting the needs of large-scale new energy access scenarios.

[0004] In the existing art, some approaches attempt to predict power transformer operating parameters through time series analysis and deep learning methods, but these methods generally suffer from insufficient modeling of periodic features or inadequate mining of multivariate correlations. For example, the patent "A Method for Predicting the Operating Status of a Power Transformer" (Patent No.: CN202410587780) proposes a deep learning-based prediction model that combines historical operating data to predict transformer status. However, this method primarily relies on feature extraction at a single time scale and fails to effectively capture the periodic patterns of variables such as transformer oil temperature and load at multiple time scales, such as daily, weekly, and seasonal. Furthermore, its projection layer typically uses simple linear transformations or fixed-order Chebyshev polynomials, which struggle to fully model the complex nonlinear dynamics of long-term periodic patterns. This results in a significant decrease in prediction accuracy for long-sequence prediction tasks. Furthermore, this model has poor adaptability to dynamic load fluctuations caused by the integration of new energy sources, limiting prediction accuracy and applicability. Another patent, "A Method for Predicting Power Transformer Oil Temperature Based on Multi-Factor Feature Fusion," integrates multiple influencing factors to predict oil temperature, attempting to leverage the relationships between multiple variables to improve prediction performance. However, this approach lacks explicit modeling of periodic features during feature fusion, and the projection layer design is relatively simple, failing to capture long-term periodic changes (such as seasonal fluctuations) through efficient nonlinear transformations. This limits the model's robustness and generalization capabilities under complex conditions. Furthermore, the projection layer of existing methods typically has a large number of parameters that scale linearly with the length of the prediction sequence, resulting in low computational efficiency and difficulty meeting the needs of large-scale real-time prediction.

[0005] Therefore, developing a prediction method that can accurately capture the periodic characteristics of transformer oil temperature and load variables and adapt to the complex working conditions of new energy has become a technical problem that needs to be solved urgently. This invention aims to propose a hybrid model-based power transformer periodic time series prediction method and a multivariate periodic time series prediction method by integrating multiple deep learning technologies and optimized projection mechanisms to improve prediction accuracy and robustness, provide technical support for energy dispatch optimization and intelligent operation and maintenance of power systems under dual carbon, and contribute to the stability of low-carbon power systems. Summary of the Invention

[0006] The purpose of the present invention is to provide a hybrid model-based power transformer periodic time series prediction method and a multivariate periodic time series prediction method. By integrating multiple deep learning technologies, the daily, weekly and seasonal periodic characteristics of power transformer oil temperature and load variables are accurately captured, and high-precision prediction sequences are generated to provide efficient technical support for energy scheduling optimization and intelligent operation and maintenance.

[0007] A method for predicting periodic time series of power transformers based on a hybrid model, characterized by comprising the following steps: Step 1: Data collection and preprocessing Collect power transformer operating data, including oil temperature and load variables; Standardize the data and handle missing values; By embedding periodic time features, a high-dimensional input feature matrix containing daily, weekly, and seasonal periodic information is generated; Step 2: Build a multi-branch hybrid model Design a multi-branch feature extraction architecture, including a bidirectional state-space encoder, a temporal convolutional network, a local window attention mechanism, and a periodic enhancement network to capture long-term periodic dependencies, multi-scale local periodic features, short-term periodic patterns, and periodic feature representations, respectively; Step 3: Feature fusion and prediction Multi-branch periodic features are fused through linear layers to generate comprehensive feature representations; A single-layer Chebyshev polynomial projection module is used to perform efficient nonlinear transformation on the fused features to generate periodic prediction sequences of oil temperature and load variables.

[0008] Preferably, the multi-branch feature extraction architecture of step 2 includes: Bidirectional State Space Encoder: Captures long-term cyclical dependencies of time series, including weekly and seasonal cycles, based on input embedding features through forward and reverse state space transformations; Temporal Convolutional Network: Based on the state-space encoder output, it uses multi-scale dilated convolution and gating mechanisms to extract multi-scale local periodic features, including diurnal cycles; Local window attention mechanism: Based on the state-space encoder output, it captures short-term periodic patterns, including sub-daily periods, by limiting the size of the attention window; Periodic Enhancement Network: Directly based on the input embedded features, it generates periodic feature enhanced representation through linear transformation and nonlinear activation.

[0009] Furthermore, in the bidirectional state space encoder, the forward and reverse state space modules are initialized, the input and output dimensions, the state dimension, the convolution kernel size and the convolution expansion factor are set, and the internal hidden layer dimension is calculated; The embedded input sequence is fed into the encoder, which passes through multiple encoding layers, each calling the forward and reverse state-space modules; In each encoding layer, the state-space module output is added to the input residual and layer normalization is performed to generate the encoder output to capture long-term periodic dependencies.

[0010] Furthermore, in the temporal convolutional network, the dilated convolution module is initialized, and the number of input channels, the total number of output channels, the dilation factor and the convolution kernel set are set; The number of output channels for each convolution kernel; The reshaped input sequence is input into each convolution kernel to generate multiple feature maps, which are spliced along the channel dimension; Generate filtered output and gated output through filtered convolution and gated convolution respectively; The input sequence is reshaped into each convolution kernel to generate a feature map, which is then concatenated along the channel dimension and processed through a linear layer to generate the temporal convolutional network output.

[0011] Furthermore, in the local window attention mechanism, the local window transformer is initialized, and the model dimension, number of attention heads, window size, number of layers, and dropout rate are set; Generate a mask tensor for the input sequence, initialized to all 1s; For each position in the sequence , set the mask positions within the preset range to 0 and convert them into Boolean tensors; In each transformer layer, the mask is applied to the multi-head attention module; The attention outputs of each head are concatenated and the final output is generated through linear transformation, residual connection and layer normalization to capture short-term periodic patterns; Connect the multi-head attention output through residual connections; Application layer normalization processing; Processed by a feedforward network, consisting of two linear layers and an activation function; Apply residual connections and layer normalization; Preferably, the feature fusion in step three includes the following steps: The output features of the bidirectional state-space encoder, temporal convolutional network, local window attention mechanism, and periodic enhancement network are concatenated along the feature dimension to generate a high-dimensional feature tensor; The concatenated high-dimensional feature tensor is input into the linear layer, and the feature dimension is reduced through linear transformation to generate reduced-dimensional features; A linear rectification activation function is applied to the dimensionality reduction features, introducing nonlinear transformation to generate the final fusion features.

[0012] Preferably, the Chebyshev polynomial projection in step 3 includes: Apply the hyperbolic tangent activation function to the fused features to limit the feature values to the range of [-1, 1]; Initialize a single-layer Chebyshev module, set the input dimension, output dimension, and polynomial order, and define the learnable coefficient matrix. The initialization value uses the standard normal distribution multiplied by 0.01 for small-scale initialization; Calculate the Chebyshev polynomial basis function and generate polynomial tensors from order 0 to order K based on the input features; Through tensor operations, the Chebyshev polynomial tensor and the learnable coefficient matrix are weighted summed to generate a prediction sequence; The output series are reshaped into the target forecast dimension to generate periodic forecast series of oil temperature and load variables, and the original dimensions are restored by denormalization.

[0013] The present invention also provides a multivariable periodic time series prediction method, using any of the above-mentioned hybrid model-based power transformer periodic time series prediction methods, comprising the following steps: Collect load variables and oil temperature data on the high-voltage, medium-voltage, and low-voltage sides of the transformer to generate multivariable input data; Input the multivariate data into the preprocessing process of step 1 in the hybrid model-based power transformer periodic time series prediction method to generate a high-dimensional input feature matrix; The multi-variable feature matrix is processed by the multi-branch hybrid model in step 2 of the hybrid model-based power transformer periodic time series prediction method to generate a periodic prediction sequence including oil temperature and load variables on each side.

[0014] The beneficial effects of the present invention include: 1) A four-branch hybrid model is proposed, which integrates a bidirectional state-space encoder, a temporal convolutional network, a local window attention mechanism, and a periodic enhancement network. Through cascade and parallel processing, it collaboratively captures long-term periodic dependencies, multi-scale local periodic features, short-term periodic patterns, and enhanced representations of periodic features, effectively modeling the complex nonlinear relationships of transformer operating data.

[0015] 2) Timestamps are encoded by sine and cosine functions, and multi-scale periodic features such as days, weeks, and seasons are explicitly embedded. The periodicity expression is further enhanced by combining a periodicity enhancement network, overcoming the limitations of traditional methods in insufficient periodic modeling.

[0016] 3) A fusion mechanism of feature splicing and linear dimensionality reduction is adopted, combined with a single-layer Chebyshev polynomial projection module. Through fixed-order design and a learnable coefficient matrix, the nonlinear modeling capability is significantly enhanced. At the same time, small-scale initialization is used to reduce computational complexity and improve parameter efficiency.

[0017] Compared with traditional methods or single models, this method significantly improves the prediction accuracy of oil temperature and load variables through four-branch collaborative modeling and multi-order projection, accurately captures multi-scale periodic characteristics, and adapts to the dynamic changes and non-stationarity of power system operation data under the scenario of new energy access. The generated high-precision periodic prediction sequence provides reliable load and oil temperature predictions for energy dispatch, supports equipment maintenance plan optimization, and improves operation and maintenance efficiency. The single-layer Chebyshev projection module reduces the number of parameters and computational complexity through fixed-order design and small-scale coefficient initialization. It is suitable for multivariate time series prediction under complex operating conditions, ensures the robustness and stability of the prediction results, and provides technical support for the stable operation of low-carbon power systems under dual carbon. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention provides the following drawings to more clearly illustrate the implementation methods and beneficial effects of the technical solution: Figure 1 This is a model diagram of the present invention, showing the overall process and multi-branch hybrid model architecture of the power transformer periodic time series prediction method based on the hybrid model.

[0019] Figure 2 FIG2 is a diagram of a bidirectional state space branch of the present invention, showing the structure and implementation of a bidirectional state space encoder.

[0020] Figure 3 This is an illustration of the time domain convolution branch of the present invention, showing the structure of the temporal convolutional network and the multi-scale feature extraction and gating mechanism.

[0021] Figure 4 This is an illustration of the local window attention branch of the present invention, showing the structure of the local window transformer and the implementation method of the local window attention mechanism. DETAILED DESCRIPTION

[0022] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings. The processes, conditions, experimental methods, etc. for implementing the present invention, except for those specifically mentioned below, are common knowledge and common common sense in the art and are not particularly limited by the present invention.

[0023] like Figure 1-4 As shown, the present invention proposes a periodic time series prediction method for power transformers and a multivariable periodic time series prediction method based on a hybrid model. By predicting oil temperature and load variables, the operating status of the transformer is evaluated to provide support for energy scheduling and intelligent operation and maintenance of the power system.

[0024] A method for predicting periodic time series of power transformers based on a hybrid model, characterized by comprising the following steps: Step 1: Data collection and preprocessing Collect power transformer operating data, including oil temperature and load variables; Standardize the data and handle missing values; By embedding periodic time features, a high-dimensional input feature matrix containing daily, weekly, and seasonal periodic information is generated; Step 2: Build a multi-branch hybrid model Design a multi-branch feature extraction architecture, including a bidirectional state-space encoder, a temporal convolutional network, a local window attention mechanism, and a periodic enhancement network to capture long-term periodic dependencies, multi-scale local periodic features, short-term periodic patterns, and periodic feature representations, respectively; Step 3: Feature fusion and prediction Multi-branch periodic features are fused through linear layers to generate comprehensive feature representations; A single-layer Chebyshev polynomial projection module is used to perform efficient nonlinear transformation on the fused features to generate periodic prediction sequences of oil temperature and load variables.

[0025] The multi-branch feature extraction architecture of step 2 includes: Bidirectional State Space Encoder: Captures long-term cyclical dependencies of time series, including weekly and seasonal cycles, based on input embedding features through forward and reverse state space transformations; Temporal Convolutional Network: Based on the state-space encoder output, it uses multi-scale dilated convolution and gating mechanisms to extract multi-scale local periodic features, including diurnal cycles; Local window attention mechanism: Based on the state-space encoder output, it captures short-term periodic patterns, including sub-daily periods, by limiting the size of the attention window; Periodic Enhancement Network: Directly based on the input embedded features, it generates periodic feature enhanced representation through linear transformation and nonlinear activation.

[0026] In the bidirectional state space encoder, the forward and reverse state space modules are initialized, the input and output dimensions, state dimensions, convolution kernel size and convolution expansion factor are set, and the internal hidden layer dimensions are calculated by the formula: in, is the input and output dimension, is the convolution expansion factor; The embedded input sequence is fed into the encoder, which passes through multiple encoding layers, each calling the forward and reverse state-space modules; In each encoding layer, the state-space module output is added to the input residual and layer normalization is performed to generate the encoder output to capture long-term periodic dependencies.

[0027] In the temporal convolutional network, the dilated convolution module is initialized, and the number of input channels, the total number of output channels, the dilation factor, and the convolution kernel set are set; The number of output channels of each convolution kernel is calculated by the formula: Among them, cout is the total number of output channels, kernel is the convolution kernel set; Initialize the two-dimensional convolution for each convolution kernel size, setting the sum size to (1, ), the expansion rate is (1, ), fill formula: in, is the convolution kernel size, is the expansion factor; The reshaped input sequence is input into each convolution kernel to generate multiple feature maps, which are spliced along the channel dimension; The filtered output and gated output are generated by filtering convolution and gated convolution respectively. The calculation formula is as follows: in, is the input sequence, is the tangent function, is the activation function; The input sequence is reshaped into each convolution kernel to generate a feature map, which is then concatenated along the channel dimension and processed through a linear layer to generate the temporal convolutional network output.

[0028] In the local window attention mechanism, initialize the local window transformer, set the model dimension, number of attention heads, window size, number of layers, and dropout rate; Generate a mask tensor for the input sequence, initialized to all 1s; For each position in the sequence , place the mask in the range Set the positions inside to 0 and convert to a Boolean tensor, where is the window size; In each transformer layer, the mask is applied to the multi-head attention module as follows: in, is the attention score matrix, is the scaling factor to prevent the value from being too large. is the mask matrix, generated based on Boolean mask; The attention outputs of each head are concatenated and the final output is generated through linear transformation, residual connection and layer normalization to capture the short-term periodic pattern. The formula is as follows: in, is the output projection matrix.

[0029] The multi-head attention output is connected through the residual connection, the formula is as follows: Application layer normalization processing, the formula is as follows: Processed through a feedforward network, it includes two linear layers and an activation function, as follows: Apply residual connection and layer normalization, the formula is as follows: The feature fusion in step 3 includes the following steps: The output features of the bidirectional state-space encoder, temporal convolutional network, local window attention mechanism, and periodic enhancement network are concatenated along the feature dimension to generate a high-dimensional feature tensor; The concatenated high-dimensional feature tensor is input into the linear layer, and the feature dimension is reduced through linear transformation to generate reduced-dimensional features; A linear rectification activation function is applied to the dimensionality reduction features, introducing nonlinear transformation to generate the final fusion features.

[0030] The Chebyshev polynomial projection in step 3 includes: Apply the hyperbolic tangent activation function to the fused features to limit the feature values to the range of [-1, 1]; Initialize a single-layer Chebyshev module, set the input dimension, output dimension, and polynomial order, and define the learnable coefficient matrix. The initialization value uses the standard normal distribution multiplied by 0.01 for small-scale initialization. The formula is as follows: Calculate the Chebyshev polynomial basis function and generate polynomial tensors from order 0 to order K based on the input features. The recursive formula is as follows: Through tensor operations, the Chebyshev polynomial tensor and the learnable coefficient matrix are weighted summed to generate a prediction sequence. The formula is as follows: in, is the input feature, For the The learnable coefficient of order, is the output prediction sequence.

[0031] Reshape the output sequence into the target prediction dimension to generate the periodic prediction sequence of oil temperature and load variables, and restore the original dimension by denormalization. The formula is as follows: in, and are the standard deviation and mean of the input data, respectively.

[0032] The present invention also provides a multivariable periodic time series prediction method, using any of the above-mentioned hybrid model-based power transformer periodic time series prediction methods, comprising the following steps: Collect load variables and oil temperature data on the high-voltage, medium-voltage, and low-voltage sides of the transformer to generate multivariable input data; Input the multivariate data into the preprocessing process of step 1 in the hybrid model-based power transformer periodic time series prediction method to generate a high-dimensional input feature matrix; The multi-variable feature matrix is processed by the multi-branch hybrid model in step 2 of the hybrid model-based power transformer periodic time series prediction method to generate a periodic prediction sequence including oil temperature and load variables on each side.

[0033] The method is implemented through the following technical solutions: First, power transformer operating data, including oil temperature and multi-level load variables, is collected and preprocessed. Preprocessing includes data normalization, linear interpolation of missing values, outlier removal, and periodic time feature embedding. Normalization eliminates dimensional differences by transforming the data to a distribution with a mean of 0 and a standard deviation of 1. Missing values are filled using linear interpolation to ensure data integrity; outliers are identified and removed using statistical methods to reduce noise. Periodic time feature embedding encodes timestamps using sine and cosine functions. Based on the timestamp, periodic features such as month, day, week, and hour are decomposed. High-dimensional time features are generated through the embedding layer and concatenated with the normalized data to form a high-dimensional input feature matrix containing multi-scale periodic information, providing rich time series information for the model.

[0034] Secondly, a multi-branch hybrid model is constructed, designing a feature extraction architecture consisting of a bidirectional state-space encoder, a temporal convolutional network, a local window attention mechanism, and a periodic enhancement network. These architectures capture the long-term periodic dependencies, multi-scale local periodic features, short-term periodic patterns, and enhanced representations of periodic features, respectively. The bidirectional state-space encoder directly processes the input embedding features through forward and reverse state-space transformations, extracting long-term trends of weekly and seasonal cycles. The temporal convolutional network, based on the state-space encoder output, employs multi-scale dilated convolutions and a gating mechanism to capture local features of daily cycles. The local window attention mechanism, based on the state-space encoder output, enhances the short-term representation of sub-daily cycles by limiting the attention window size and implementing a high-resolution sharding strategy. The periodic enhancement network directly processes the input embedding features and generates enhanced representations of periodic features through linear transformations and nonlinear activations. Each branch processes the input feature matrix in parallel or in cascade to generate diverse periodic feature representations.

[0035] Next, the multi-branch features are integrated through the feature fusion module. This module concatenates the output features of the bidirectional state-space encoder, temporal convolutional network, local window attention mechanism, and periodic reinforcement network along the feature dimension to generate a high-dimensional feature tensor. This is then reduced in dimension through a linear layer and a nonlinear activation function is applied to generate a comprehensive feature representation. The fused multi-branch features are then fed into an efficient single-layer Chebyshev polynomial projection module for nonlinear transformation, generating a periodic prediction series for the oil temperature and load variables.

[0036] By integrating multiple deep learning techniques, including a state-space encoder, a temporal convolutional network, a local window attention mechanism, and a periodic enhancement network, this paper significantly improves the ability to model periodic features at multiple time scales, such as daily, weekly, and seasonal. The state-space encoder captures long-term periodic dependencies based on a state-space module, the temporal convolutional network extracts local periodic features through multi-scale dilated convolution, the local window attention mechanism enhances the expression of short-term periodic patterns by limiting the size of the attention window, and the periodic enhancement network strengthens the representation of periodic features through a simple fully connected layer. The present invention innovatively introduces a single-layer Chebyshev polynomial projection module, which significantly enhances the ability to model long-term periodic patterns through a fixed-order nonlinear transformation and small-scale coefficient initialization, while reducing the number of parameters and improving computational efficiency. Feature fusion uses feature concatenation and linear dimensionality reduction mechanisms to effectively integrate multi-branch features. Compared with existing technologies, the present invention has significant technical advantages in multi-scale feature decomposition, feature fusion efficiency, and projection layer optimization, providing an efficient solution for power transformer state prediction in new energy scenarios.

[0037] This implementation uses the ETTm1 dataset as the experimental data source to verify the effectiveness of a hybrid model-based method for predicting power transformer oil temperature and load variables. The ETTm1 dataset is a multivariate time series dataset that records key parameters during transformer operation, including oil temperature and related load variables. It is suitable for multivariate prediction of multiple variables. By predicting oil temperature and load variables, this method can assess the operating status and potential failure risks of transformers, providing support for intelligent monitoring and maintenance of power systems. The following details the dataset collection, partitioning, and preprocessing process.

[0038] The ETTm1 dataset, collected in real time by sensors and monitoring systems on power transformers, contains multiple time series variables with complex nonlinear relationships and dynamic characteristics, such as the impact of sudden load changes or ambient temperature fluctuations on oil temperature. This experiment uses the ETTm1 subset as an example. This subset is sampled at a 15-minute interval and contains approximately 69,680 data points (2 years × 365 days × 24 hours × 4 times / hour).

[0039] The following describes the implementation steps of the technical solution in detail with reference to specific embodiments: Step 1: Based on the requirements of the prediction task, the ETTm1 dataset is divided into a training set, a validation set, and a test set to support model training, parameter adjustment, and performance evaluation. The specific division method is as follows: Training set: data from the first 12 months, used for initial learning of model parameters.

[0040] Validation set: The next 4 months of data, used for hyperparameter tuning and early stopping mechanism.

[0041] Test set: Data from the last 8 months, used to evaluate the model's prediction performance and failure risk assessment capabilities.

[0042] Step 2: Define the input sequence length, label length, and prediction length. In this experiment, the sequence length is set to 96 time steps, the label length is 48, and the prediction length is 96 time steps. Using a sliding window approach, continuous input and output sequence pairs are extracted from the dataset. Each window contains 96 time steps of input data and 96 time steps of prediction targets. Time tags are extracted from the timestamps to provide periodicity characteristics.

[0043] Step 3: Use the data loader to batch load data using a sliding window approach to ensure data order and integrity. The batch size of the data loader is set based on memory capacity to avoid random shuffling and preserve temporal dependencies. Extract a sequence of consecutive time steps from the dataset as the input sequence, and extract the predicted output sequence starting from the next time step of the input sequence.

[0044] Step 4: For any missing values that may exist, use the forward interpolation method to fill them in, copying the previous non-missing value to the subsequent missing point to ensure the continuity of the time series. When the numerical ranges of different features vary greatly, model training may become unstable and convergence may slow down. In addition, there are dimensional differences between different features that have not been standardized, and outliers or noise may interfere with model training. In order to enable the model to treat each feature more fairly, ensure that different features have the same scale, improve the model's training efficiency, prediction performance, and better capture the complex relationships between features, the data needs to be standardized.

[0045] Transform the data along the time dimension to a distribution with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between different features and improving the convergence speed and prediction performance of the model.

[0046] Mean and standard deviation calculation: Calculate the mean μ and standard deviation stdev of each variable along the time dimension; Normalize the data using the calculated mean and standard deviation using the following formula: , in, is the sample size, It is The value of the sample, is the mean, is the standard deviation, is the standardized value; Step 5: Use sine and cosine functions to encode the date and minute extracted from the timestamp, embedding periodic information. The encoded time features are concatenated with the standardized operational data along the feature dimension to form the input feature matrix. The concatenated features are mapped to a high-dimensional space via an embedding layer, allowing the model to capture complex feature relationships. The final features are organized into a time series by time step and fed into the subsequent feature extraction module using the following formula: Step 6: The present invention adopts a four-branch architecture to extract time series features. The pre-processed data is processed by four parallel branches, namely, a bidirectional state space branch, a temporal convolution branch, a local window transformation branch, and a periodic enhancement branch (see Figure 1 ). The specific implementation of each branch is as follows: 1) Bidirectional state space branch Bidirectional state space branches are used to capture long-term periodic dependencies, and their structure is as follows Figure 2 As shown. The embedded feature matrix enters the forward and reverse state space transformations respectively. The forward transformation calculates the forward dependency, and the reverse transformation calculates the backward dependency by flipping the feature matrix along the time dimension. The formula is as follows: in, is the characteristic of the current variable, is the input feature, is hidden state, 、 、 , D is the training parameter; The outputs of the forward and reverse transforms are directly added to generate a composite feature. To capture global features and long-term dependencies, a multi-layer stacking process is employed, with residual connections and normalization layers added between each layer to enhance model stability. After processing, the features enter the feedforward network, where they are refined using a linear rectification activation function and finally compressed back to their original dimensions through a linear layer. This branch effectively captures long-term trends in weekly and seasonal cycles, such as seasonal fluctuations in load variables.

[0047] 2) Temporal convolution branch The temporal convolution branch extracts local periodic features through multi-scale dilated convolution and gating mechanism. Its structure is as follows: Figure 3 As shown in Figure 2, the input features are first adapted to the 2D convolution operation by dimensional rearrangement and expansion. Multi-scale dilated convolution is used to extract features in the filtering path and the gating path respectively. The convolution kernel size is 3, and the formula is as follows: in, represents the convolution kernel weight, Represents the input signal at time step The value at is the convolution kernel size, is the expansion factor.

[0048] The tensor after the convolution output is padded to align the output dimension with the input to avoid loss of boundary information. The formula is as follows: The filtering path generates features through hyperbolic tangent activation, and the gating path generates gating signals through linear activation. The two are multiplied element-wise to generate comprehensive features. This branch effectively captures local features such as diurnal cycles.

[0049] 3) Local Window Attention Branch The local window transformation branch enhances the short-term feature expression by limiting the attention range. Its structure is as follows Figure 4 As shown in Figure 2, the input data is combined with a high-resolution sharding strategy to generate a high-resolution time series. The sharded data is mapped to a high-dimensional representation through an embedding layer, and position embedding is added.

[0050] The data enters the multi-layer encoding layer, and the short-term dependency is calculated based on the local window attention mechanism. The window size is set to 24, and each token only pays attention to the surrounding tokens within the window. The attention calculation formula is: in, 、 、 are the query, key, and value vectors within the window, is the dimension of the key.

[0051] Residual connections and normalization layers are used between layers to enhance stability. This branch effectively captures short-term patterns of sub-daily cycles.

[0052] 4) Periodic Enhancement Branch The periodic enhancement branch directly processes the embedded features and enhances the periodic features through linear transformation and nonlinear activation. The features are processed through a two-layer linear network and linear rectification activation.

[0053] Step 7: Integrate multi-branch features through the feature fusion module to generate a comprehensive feature representation. The specific steps are as follows: The four branches are concatenated along the feature dimension to generate a high-dimensional feature tensor.

[0054] Step 8: Input the fused features into the single-layer Chebyshev polynomial to generate the prediction sequence. This module is implemented by the following steps: Apply the hyperbolic tangent activation function to limit the eigenvalues to the range [-1, 1]; Initializes a single-layer Chebyshev module, sets the input dimension, output dimension, and fixed polynomial order, and defines the learnable coefficient matrix. The coefficient matrix is initialized using a small-scale normal distribution to ensure training stability.

[0055] Computes Chebyshev polynomial basis functions of order 0 to K, generating a polynomial tensor.

[0056] Through tensor operations, the Chebyshev polynomial tensor and the learnable coefficient matrix are weighted summed to generate a prediction sequence.

[0057] The output series is reshaped into the target forecast dimension to generate the periodic forecast series of oil temperature and load variables, and the original dimension is restored through the denormalization operation.

[0058] To validate the real-world application of this invention, we evaluated the proposed hybrid model on four real-world datasets (ETTH1, ETTH2, ETTm1, and ETTm2) and compared it with five multivariate time series forecasting models from the past three years. The experimental results are summarized in Table 1.

[0059] Table 1 Experimental results As can be seen from Table 1, compared with traditional methods or single models, this method significantly improves the prediction accuracy of oil temperature and load variables through four-branch collaborative modeling and multi-order projection, accurately captures multi-scale periodic characteristics, and adapts to the dynamic changes and non-stationarity of power system operation data under the scenario of new energy access. The generated high-precision periodic prediction series provides reliable load and oil temperature predictions for energy dispatch, supports equipment maintenance plan optimization, and improves operation and maintenance efficiency. The single-layer Chebyshev projection module reduces the number of parameters and computational complexity through fixed-order design and small-scale coefficient initialization. It is suitable for multivariate time series prediction under complex operating conditions, ensures the robustness and stability of the prediction results, and provides technical support for the stable operation of low-carbon power systems under dual carbon.

Claims

1. A method for predicting periodic time series of power transformers based on a hybrid model, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing Collect power transformer operating data, including oil temperature and load variables; Standardize the data and handle missing values; By embedding periodic time features, a high-dimensional input feature matrix containing daily, weekly, and seasonal periodic information is generated; Step 2: Build a multi-branch hybrid model Design a multi-branch feature extraction architecture, including a bidirectional state-space encoder, a temporal convolutional network, a local window attention mechanism, and a periodic enhancement network to capture long-term periodic dependencies, multi-scale local periodic features, short-term periodic patterns, and periodic feature representations, respectively; Step 3 : Feature fusion and prediction Multi-branch periodic features are fused through linear layers to generate comprehensive feature representations; A single-layer Chebyshev polynomial projection module is used to perform efficient nonlinear transformation on the fused features to generate periodic prediction sequences of oil temperature and load variables.

2. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 1, characterized in that: The multi-branch feature extraction architecture of step 2 includes: Bidirectional State Space Encoder: Captures long-term cyclical dependencies of time series, including weekly and seasonal cycles, based on input embedding features through forward and reverse state space transformations; Temporal Convolutional Network: Based on the state-space encoder output, it uses multi-scale dilated convolution and gating mechanisms to extract multi-scale local periodic features, including diurnal cycles; Local window attention mechanism: Based on the state-space encoder output, it captures short-term periodic patterns, including sub-daily periods, by limiting the size of the attention window; Periodic Enhancement Network: Directly based on the input embedded features, it generates periodic feature enhanced representation through linear transformation and nonlinear activation.

3. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 2, characterized in that: In the bidirectional state space encoder, the forward and reverse state space modules are initialized, the input and output dimensions, the state dimension, the convolution kernel size and the convolution expansion factor are set, and the internal hidden layer dimension is calculated; The embedded input sequence is fed into the encoder, which passes through multiple encoding layers, each calling the forward and reverse state-space modules; In each encoding layer, the state-space module output is added to the input residual and layer normalization is performed to generate the encoder output to capture long-term periodic dependencies.

4. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 2, characterized in that: In the temporal convolutional network, the dilated convolution module is initialized, and the number of input channels, the total number of output channels, the dilation factor, and the convolution kernel set are set; Calculate the number of output channels of each convolution kernel; The reshaped input sequence is input into each convolution kernel to generate multiple feature maps, which are spliced along the channel dimension; Generate filtered output and gated output through filtered convolution and gated convolution respectively; The input sequence is reshaped into each convolution kernel to generate a feature map, which is then concatenated along the channel dimension and processed through a linear layer to generate the temporal convolutional network output.

5. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 2, characterized in that: In the local window attention mechanism, initialize the local window transformer, set the model dimension, number of attention heads, window size, number of layers, and dropout rate; Generate a mask tensor for the input sequence, initialized to all 1s; For each position in the sequence , set the mask positions within the preset range to 0 and convert them into Boolean tensors; In each transformer layer, the mask is applied to the multi-head attention module; The attention outputs of each head are concatenated and the final output is generated through linear transformation, residual connection and layer normalization to capture short-term periodic patterns; Connect the multi-head attention output through residual connections; Application layer normalization processing; Processed by a feedforward network, consisting of two linear layers and an activation function; Apply residual connections and layer normalization.

6. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 1, characterized in that: The feature fusion in step 3 includes the following steps: The output features of the bidirectional state-space encoder, temporal convolutional network, local window attention mechanism, and periodic enhancement network are concatenated along the feature dimension to generate a high-dimensional feature tensor; The concatenated high-dimensional feature tensor is input into the linear layer, and the feature dimension is reduced through linear transformation to generate reduced-dimensional features; A linear rectification activation function is applied to the dimensionality reduction features, introducing nonlinear transformation to generate the final fusion features.

7. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 1, characterized in that: The Chebyshev polynomial projection in step 3 includes: Apply the hyperbolic tangent activation function to the fused features to limit the feature values to the range of [-1, 1]; Initialize a single-layer Chebyshev module, set the input dimension, output dimension, and polynomial order, and define the learnable coefficient matrix. The initialization value uses the standard normal distribution multiplied by 0.01 for small-scale initialization; Calculate the Chebyshev polynomial basis function and generate polynomial tensors from order 0 to order K based on the input features; Through tensor operations, the Chebyshev polynomial tensor and the learnable coefficient matrix are weighted summed to generate a prediction sequence; The output series are reshaped into the target forecast dimension to generate periodic forecast series of oil temperature and load variables, and the original dimensions are restored by denormalization.

8. A multivariate periodic time series prediction method, using the hybrid model-based power transformer periodic time series prediction method according to any one of claims 1 to 7, characterized in that: The following steps are involved: Collect load variables and oil temperature data on the high-voltage, medium-voltage, and low-voltage sides of the transformer to generate multivariable input data; Input the multivariate data into the preprocessing process of step 1 in the hybrid model-based power transformer periodic time series prediction method to generate a high-dimensional input feature matrix; The multi-variable feature matrix is processed by the multi-branch hybrid model in step 2 of the hybrid model-based power transformer periodic time series prediction method to generate a periodic prediction sequence including oil temperature and load variables on each side.

Images