Hybrid model-based power transformer periodic time series prediction method and multivariate periodic time series prediction method

Abstract

The application discloses a power transformer periodic time series prediction method and a multivariate periodic time series prediction method based on a mixed model. The method collects transformer operation data, including oil temperature and load variables, carries out standardization, missing value processing and periodic time feature embedding, constructs a multi-branch mixed model, fuses a bidirectional state space encoder, a time convolution network, a local window attention mechanism and a periodic enhancement network, respectively captures long-term periodic dependence, multi-scale local periodic characteristics, short-term periodic patterns and periodic feature enhancement representation, adaptively integrates multi-branch features through a global context weighted fusion mechanism, adopts a multi-order Chebyshev polynomial projection module to carry out efficient nonlinear transformation on the fused features, and generates a prediction sequence of the oil temperature and the load variable.

Description

TECHNICAL FIELD

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

[0002] The energy structure in China is accelerating the transformation to clean and low carbon, which requires the power system to significantly increase the proportion of new energy (such as wind energy and solar energy) access, and to optimize energy utilization efficiency to reduce carbon emissions and ensure energy security. However, new energy generation has the characteristics of intermittency and volatility, which leads to high dynamicity and periodicity of power flow direction and load change in the power system, and puts forward higher requirements for the operation stability of power transformers and the efficiency of energy scheduling.

[0003] As the core equipment of the power system, the oil temperature and load variables (such as high-voltage side load, medium-voltage side load, and low-voltage side load) of the power transformer usually show obvious daily, weekly, and seasonal periodic characteristics. These periodic patterns directly affect power scheduling, load balancing, and equipment maintenance efficiency. Accurate prediction of transformer oil temperature and load variables not only helps to optimize new energy consumption and power scheduling, but also reduces energy consumption and carbon emissions through advance planning of equipment maintenance. However, traditional transformer operating parameter prediction methods mainly rely on simple statistical models or artificial experience, which are 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 prior art, some schemes try to predict the operation parameters of power transformers through time series analysis and deep learning methods, but there are generally problems of insufficient modeling of periodic characteristics or insufficient mining of multivariate correlations. For example, the patent "Power transformer operation state prediction method" (Patent No. CN202410587780) proposes a prediction model based on deep learning, which combines historical operation data to predict the state of the transformer. However, this method mainly relies on single time scale feature extraction and fails to effectively capture the periodicity of transformer oil temperature and load variables at multiple time scales such as day, week, and season. Moreover, its projection layer usually uses simple linear transformation or fixed-order Chebyshev polynomials, which cannot adequately model the complex nonlinear dynamics of long-term periodic patterns, resulting in a significant decrease in prediction accuracy in long sequence prediction tasks. In addition, its model has poor adaptability to dynamic load fluctuations caused by new energy access, limiting the prediction accuracy and applicability. Another patent "Power transformer oil temperature prediction method based on multi-factor feature fusion" attempts to improve prediction accuracy by utilizing the relationship between multiple variables. However, this method lacks explicit modeling of periodic characteristics 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, resulting in limited robustness and generalization ability of the model under complex conditions. At the same time, the projection layer of existing methods usually has a large number of parameters, which increases linearly with the length of the prediction sequence, resulting in low computational efficiency and difficulty in meeting large-scale real-time prediction requirements.

[0005] Therefore, it is a technical problem to be solved to develop a prediction method that can accurately capture the periodic characteristics of transformer oil temperature and load variables and adapt to complex conditions of new energy. The present 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 techniques 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, and contribute to the stability of low-carbon power systems. SUMMARY

[0006] The present invention aims to provide a hybrid model-based power transformer periodic time series prediction method and a multivariate periodic time series prediction method, which accurately capture the daily, weekly, and seasonal periodic characteristics of power transformer oil temperature and load variables by integrating multiple deep learning techniques, generating high-precision prediction sequences, and providing efficient technical support for energy dispatch optimization and intelligent operation and maintenance.

[0007] A hybrid model-based power transformer periodic time series prediction method, characterized by the following steps:

[0008] Step 1: Data collection and preprocessing

[0009] Collecting power transformer operation data, including oil temperature and load variables;

[0010] Standardizing and processing missing values of the data;

[0011] Generating a high-dimensional input feature matrix containing daily, weekly, and seasonal cycle information through periodic time feature embedding;

[0012] Step two: building a multi-branch hybrid model

[0013] Designing a multi-branch feature extraction architecture, including a bidirectional state space encoder, a time 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;

[0014] Step three: feature fusion and prediction

[0015] Fusing multi-branch periodic features through a linear layer to generate a comprehensive feature representation;

[0016] Using a single-layer Chebyshev polynomial projection module to perform nonlinear transformation on the fused features to generate periodic prediction sequences for oil temperature and load variables.

[0017] Preferably, the multi-branch feature extraction architecture of step two includes:

[0018] Bidirectional state space encoder: capturing long-term periodic dependencies of time series based on input embedding features through forward and backward state space transformations, including weekly and seasonal cycles;

[0019] Time convolutional network: based on the output of the state space encoder, using multi-scale dilated convolution and gating mechanism to extract multi-scale local periodic features, including daily cycles;

[0020] Local window attention mechanism: based on the output of the state space encoder, capturing short-term periodic patterns by limiting the attention window size, including sub-daily cycles;

[0021] Periodic enhancement network: directly based on input embedding features, generating periodic feature enhancement representations through linear transformation and nonlinear activation.

[0022] Further, in the bidirectional state space encoder, initialize the forward and backward state space modules, set the input and output dimensions, state dimensions, convolution kernel size, and convolution expansion factor, and calculate the internal hidden layer dimension through;

[0023] Input the embedded input sequence into the encoder, and call the forward and backward state space modules through multiple encoding layers;

[0024] In each encoding layer, the state space module outputs are added to the input residuals, and the encoder outputs are generated by layer normalization to capture long-term periodic dependencies.

[0025] Further, in the time convolution network, the dilated convolution module is initialized, the input channel number, the total output channel number, the dilation factor and the convolution kernel set are set;

[0026] The output channel number of each convolution kernel;

[0027] The reshaped input sequence is input into each convolution kernel to generate multiple feature maps, which are spliced along the channel dimension;

[0028] The filter output and the gate output are generated by filter convolution and gate convolution respectively;

[0029] The input sequence is reshaped and input into each convolution kernel to generate feature maps, which are spliced along the channel dimension and processed by a linear layer to generate the output of the time convolution network.

[0030] Further, in the local window attention mechanism, the local window transformer is initialized, and the model dimension, the number of attention heads, the window size, the number of layers, and the dropout rate are set;

[0031] A mask tensor is generated for the input sequence and is initialized to all 1s;

[0032] For each position in the sequence The positions of the mask within the preset range are set to 0, and the mask is converted into a Boolean tensor;

[0033] In each layer of the transformer, the mask is applied to the multi-head attention module;

[0034] The outputs of each head of attention are spliced, processed by linear transformation, residual connection and layer normalization to generate the final output, and short-term periodic patterns are captured;

[0035] The multi-head attention output is processed by residual connection;

[0036] Layer normalization is applied;

[0037] The output is processed by a feedforward network, which includes two linear layers and an activation function;

[0038] Residual connection and layer normalization are applied;

[0039] Preferably, the feature fusion in step three includes the following steps:

[0040] The output features of the bidirectional state space encoder, the time convolution network, the local window attention mechanism and the periodic enhancement network are spliced along the feature dimension to generate a high-dimensional feature tensor;

[0041] The concatenated high-dimensional feature tensor is input into a linear layer, and the feature dimension is reduced through linear transformation to generate a reduced dimension feature;

[0042] A linear rectifier activation function is applied to the reduced dimension feature to introduce a nonlinear transformation to generate a final fusion feature.

[0043] Preferably, the Chebyshev polynomial projection in step three comprises:

[0044] A hyperbolic tangent activation function is applied to the fusion feature to limit the feature value to the range [-1, 1];

[0045] A single-layer Chebyshev module is initialized, the input dimension, output dimension and polynomial order are set, and the learnable coefficient matrix is defined, and the standard normal distribution multiplied by 0.01 is used for small-scale initialization of the initialization value;

[0046] The Chebyshev polynomial basis function is calculated, and a polynomial tensor from 0 order to K order is generated based on the input feature;

[0047] The Chebyshev polynomial tensor and the learnable coefficient matrix are weighted and summed through tensor operation to generate a prediction sequence;

[0048] The output sequence is reshaped to the target prediction dimension to generate a periodic prediction sequence of the oil temperature and load variable, and the original dimension is restored through inverse normalization.

[0049] The application also provides a multivariate periodic time series prediction method, which uses any one of the above-mentioned power transformer periodic time series prediction methods based on a hybrid model, comprising the following steps:

[0050] The load variable and oil temperature data of the high-voltage side, medium-voltage side and low-voltage side of the transformer are collected to generate multivariate input data;

[0051] The multivariate data is input into the preprocessing process of step one of the power transformer periodic time series prediction method based on a hybrid model to generate a high-dimensional input feature matrix;

[0052] The multivariate feature matrix is processed by the multi-branch hybrid model of step two of the power transformer periodic time series prediction method based on a hybrid model to generate a periodic prediction sequence containing the oil temperature and the load variable of each side.

[0053] The application has the following advantages:

[0054] 1) A four-branch hybrid model that combines bidirectional state space encoder, time convolution network, local window attention mechanism and periodic enhancement network is proposed. Through cascading and parallel processing, long-term periodic dependence, multi-scale local periodic features, short-term periodic patterns and periodic feature enhancement representation are captured to effectively model the complex nonlinear relationship of transformer operation data.

[0055] 2) The timestamp is encoded by sine and cosine functions to explicitly embed multi-scale periodic features such as day, week and season. Combined with the periodic enhancement network, the periodic expression is further strengthened to overcome the limitations of traditional periodic modeling.

[0056] 3) A feature splicing and linear dimension reduction fusion mechanism is adopted, combined with a single-layer Chebyshev polynomial projection module. Through fixed order design and learnable coefficient matrix, the non-linear modeling capability is significantly enhanced, while the calculation complexity is reduced and the parameter efficiency is improved through small-scale initialization.

[0057] 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 features, and adapts to the dynamic changes and non-stationarity of power system operation data under new energy access scenarios. The generated high-precision periodic prediction sequence provides reliable load and oil temperature prediction for energy scheduling, supports equipment maintenance plan optimization, and improves operational efficiency. The single-layer Chebyshev projection module reduces the parameter quantity and calculation complexity through fixed order design and small-scale coefficient initialization, which is suitable for multivariate time series prediction under complex operating conditions, ensuring the robustness and stability of the prediction results, and providing technical support for the stable operation of low-carbon power systems. BRIEF DESCRIPTION OF DRAWINGS

[0058] The present application provides the following drawings to more clearly illustrate the embodiments and beneficial effects of the technical solutions:

[0059] Figure 1 The model diagram of the present application shows the overall process of the power transformer periodic time series prediction method based on the hybrid model and the multi-branch hybrid model architecture.

[0060] Figure 2 The bidirectional state space branch diagram of the present application shows the structure and implementation of the bidirectional state space encoder.

[0061] Figure 3 The time domain convolution branch diagram of the present application shows the structure of the time convolution network and the multi-scale feature extraction and gating mechanism.

[0062] Figure 4 The local window attention branch diagram of the present application shows the structure of the local window transformer and the implementation of the local window attention mechanism. DETAILED DESCRIPTION

[0063] 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.

[0064] like Figures 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.

[0065] A method for predicting periodic time series of power transformers based on a hybrid model, characterized by comprising the following steps:

[0066] Step 1: Data collection and preprocessing

[0067] Collect power transformer operating data, including oil temperature and load variables;

[0068] Standardize the data and handle missing values;

[0069] By embedding periodic time features, a high-dimensional input feature matrix containing daily, weekly, and seasonal periodic information is generated;

[0070] Step 2: Build a multi-branch hybrid model

[0071] 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;

[0072] Step 3: Feature Fusion and Prediction

[0073] Multi-branch periodic features are fused through linear layers to generate comprehensive feature representations;

[0074] 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.

[0075] The multi-branch feature extraction architecture of step 2 includes:

[0076] 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;

[0077] Temporal convolution network: based on the state space encoder output, using multi-scale dilated convolution and gating mechanism, extracting multi-scale local periodic features, including daily periodicity;

[0078] Local window attention mechanism: based on the state space encoder output, by limiting the attention window size, capturing short-term periodic patterns, including sub-daily periodicity;

[0079] Periodic enhancement network: directly based on the input embedding features, through linear transformation and nonlinear activation, generating periodic feature enhancement representation.

[0080] 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 dimension is set through the formula:

[0081]

[0082] wherein, is the input and output dimension, is the convolution expansion factor;

[0083] The embedding input sequence is input into the encoder, and the forward and reverse state space modules are called in each encoding layer;

[0084] In each encoding layer, the state space module output is added with the input residual, and the encoder output is generated through layer normalization to capture long-term periodic dependence.

[0085] In the temporal convolution network, the dilated convolution module is initialized, the input channel number, total output channel number, dilation factor and convolution kernel set are set;

[0086] The output channel number of each convolution kernel is set through the formula:

[0087]

[0088] wherein, cout is the total output channel number, and kernel is the convolution kernel set;

[0089] A two-dimensional convolution is initialized for each convolution kernel size, and the size of and is set to (1, ), the dilated rate is (1, ), and the padding formula is:

[0090]

[0091] wherein, is the convolution kernel size, is the dilation factor;

[0092] The reshaped input sequence is input into each convolution kernel to generate multiple feature maps, which are spliced along the channel dimension;

[0093] A filtering output and a gating output are generated by filtering convolution and gating convolution respectively, and the calculation formula is as follows:

[0094]

[0095] wherein, is an input sequence, is a tangent function, is an activation function;

[0096] The input sequence is reshaped and input into each convolution kernel to generate a feature map, which is spliced along the channel dimension and processed by a linear layer to generate a time convolution network output.

[0097] In the local window attention mechanism, a local window transformer is initialized, and the model dimension, the number of attention heads, the window size, the number of layers, and the dropout rate are set.

[0098] A mask tensor is generated for the input sequence and is initialized as all 1s.

[0099] For each position in the sequence , the positions of the mask in the range are set to 0, and are converted into a Boolean tensor, wherein is the window size;

[0100] In each layer transformer, the mask is applied to the multi-head attention module, and the formula is as follows:

[0101]

[0102] wherein, is an attention score matrix, is a scaling factor to prevent large values, is a mask matrix generated based on the Boolean mask;

[0103] The outputs of each head attention are spliced, and the final output is generated by linear transformation, residual connection and layer normalization to capture short-term periodic patterns, and the formula is as follows:

[0104]

[0105] wherein, is an output projection matrix.

[0106] The multi-head attention output is connected through a residual connection, and the formula is as follows:

[0107]

[0108] Layer normalization processing is applied, and the formula is as follows:

[0109]

[0110] Processed through a feedforward network, it includes two linear layers and an activation function, as follows:

[0111]

[0112] Apply residual connection and layer normalization, the formula is as follows:

[0113]

[0114] The feature fusion in step 3 includes the following steps:

[0115] 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;

[0116] 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;

[0117] A linear rectification activation function is applied to the dimensionality reduction features, introducing nonlinear transformation to generate the final fusion features.

[0118] The Chebyshev polynomial projection in step 3 includes:

[0119] Apply the hyperbolic tangent activation function to the fused features to limit the feature values ​​to the range of [-1, 1];

[0120] 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:

[0121]

[0122] 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:

[0123]

[0124] 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:

[0125]

[0126] in, is the input feature, For the The learnable coefficient of order, to output the prediction sequence.

[0127] The output sequence is reshaped to the target prediction dimension, and the periodic prediction sequence of the oil temperature and the load variable is generated, and the original dimension is recovered through inverse normalization, and the formula is as follows:

[0128]

[0129] wherein, and are the standard deviation and the mean of the input data respectively.

[0130] The application also provides a multivariate periodic time series prediction method, which uses the power transformer periodic time series prediction method based on the hybrid model according to any one of the above, and comprises the following steps:

[0131] Collecting the load variable and the oil temperature data of the high-voltage side, the medium-voltage side and the low-voltage side of the transformer to generate multivariate input data;

[0132] Inputting the multivariate data into the preprocessing procedure of step one of the power transformer periodic time series prediction method based on the hybrid model to generate a high-dimensional input feature matrix;

[0133] Processing the multivariate feature matrix through the multi-branch hybrid model of step two of the power transformer periodic time series prediction method based on the hybrid model to generate a periodic prediction sequence containing the oil temperature and the load variable of each side.

[0134] The method is realized through the following technical solutions:

[0135] Firstly, the operation data of the power transformer is collected, including the oil temperature and the multistage load variable, and data preprocessing is performed. The preprocessing includes data normalization, linear interpolation of missing values, removal of outliers and periodic time feature embedding. Normalization transforms the data to a distribution with a mean of 0 and a standard deviation of 1, eliminating the difference in dimension. Linear interpolation is used to fill in the missing values to ensure data integrity; outliers are identified and removed by statistical methods to reduce noise interference. Periodic time feature embedding encodes the timestamp using sine and cosine functions. Based on the timestamp, the monthly, daily, weekly and hourly periodic features are decomposed, and a high-dimensional time feature is generated through an embedding layer. The normalized data is spliced to form a high-dimensional input feature matrix containing multi-scale periodic information, providing rich time series information for the model.

[0136] Secondly, a multi-branch hybrid model is constructed, and a feature extraction architecture containing a bidirectional state space encoder, a time convolution network, a local window attention mechanism and a periodic enhancement network is designed to capture long-term periodic dependence, multi-scale local periodic features, short-term periodic patterns and periodic feature enhancement representation of time series respectively. The bidirectional state space encoder directly processes the input embedding features through forward and backward state space transformation to extract the long-term trend of weekly and seasonal cycles; the time convolution network based on the state space encoder output adopts multi-scale dilated convolution and gating mechanism to capture local features of the daily cycle; the local window attention mechanism based on the state space encoder output enhances the short-term feature expression ability of the sub-daily cycle through limiting the attention window size and high-resolution slicing strategy; the periodic enhancement network directly processes the input embedding features to generate periodic feature enhancement representation through linear transformation and nonlinear activation. Each branch processes the input feature matrix in parallel or cascaded to generate diversified periodic feature representation.

[0137] Thirdly, the multi-branch features are integrated through a feature fusion module. The feature fusion module splices the output features of the bidirectional state space encoder, the time convolution network, the local window attention mechanism and the periodic enhancement network along the feature dimension to generate a high-dimensional feature tensor, and then reduces the dimension through a linear layer and applies a nonlinear activation function to generate a comprehensive feature representation. The fused multi-branch features enter the efficient single-layer Chebyshev polynomial projection module for nonlinear transformation to generate the periodic prediction sequence of the oil temperature and load variables.

[0138] The present application significantly improves the modeling ability of multi-time scale periodic features such as daily, weekly and seasonal cycles by fusing state space encoders, time convolution networks, local window attention mechanisms and periodic enhancement networks and other deep learning technologies. The state space encoder captures long-term periodic dependence based on the state space module, the time convolution network extracts local periodic features through multi-scale dilated convolution, the local window attention mechanism enhances short-term periodic pattern expression by limiting the attention window size, and the periodic enhancement network strengthens periodic feature representation through a simple fully connected layer. The present application innovatively introduces a single-layer Chebyshev polynomial projection module, which significantly enhances the modeling ability of long-term periodic patterns through fixed-order nonlinear transformation and small-scale coefficient initialization, while reducing the parameter amount and improving the computational efficiency. The feature fusion adopts feature splicing and linear dimension reduction mechanism to effectively integrate multi-branch features. Compared with the prior art, the present application has significant technical advantages in multi-scale feature decomposition, feature fusion efficiency and projection layer optimization, and provides an efficient solution for power transformer state prediction in new energy scenarios.

[0139] This embodiment takes the ETTm1 dataset as the experimental data source to verify the effectiveness of the power transformer oil temperature and load variable prediction method based on the mixed model. The ETTm1 dataset is a multivariate time series dataset that records key parameters during transformer operation, including oil temperature and related load variables, suitable for multivariate prediction tasks. Through the prediction of oil temperature and load variables, this method can assess the operating status and potential failure risk of the transformer, providing support for intelligent monitoring and maintenance of the power system. The following describes the data collection, division, and preprocessing process in detail.

[0140] The ETTm1 dataset is collected in real-time by sensors and monitoring systems on power transformers, containing multiple time series variables with complex nonlinear relationships and dynamic characteristics, such as the impact of load surges or environmental temperature fluctuations on oil temperature. This experiment takes the ETTm1 subset as an example, which has a sampling interval of 15 minutes and contains approximately 69680 data points (2 years x 365 days x 24 hours x 4 times / hour).

[0141] The following describes the implementation steps of the technical solution in detail combined with specific examples:

[0142] Step one: According to the requirements of the prediction task, divide the ETTm1 dataset into training set, validation set and test set to support model training, parameter tuning and performance evaluation. The specific division method is as follows:

[0143] Training set: The first 12 months of data, used for initial learning of model parameters.

[0144] Validation set: The next 4 months of data, used for hyperparameter tuning and early stopping mechanism.

[0145] Test set: The last 8 months of data, used to evaluate the prediction performance of the model and the ability of fault risk assessment.

[0146] Step two: 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 set to 48, and the prediction length is set to 96 time steps. By using the sliding window method, continuous input-output sequence pairs are extracted from the dataset, with each window containing 96 time steps of input data and 96 time steps of prediction targets. Time markers are extracted from the timestamp to provide periodic features.

[0147] Step three: Use the data loader to load data in batches using the sliding window method to ensure the order and integrity of the data. The batch size of the data loader is set according to the memory size to avoid random shuffling to maintain the time dependency. Continuous time step sequences are extracted from the dataset as input sequences, and prediction output sequences are extracted from the next time step of the input sequences.

[0148] Step four: For the missing values that may exist, the forward interpolation method is used to fill in the missing values, and the previous non-missing value is copied to the subsequent missing point to ensure the continuity of the time series. When the numerical range of different features is different, the training of the model may become unstable, the convergence speed becomes slow, and there is a dimensional difference between different features that have not been standardized, and abnormal values or noise will interfere with the training of the model. In order to make the model treat each feature more fairly, ensure that different features have the same scale, improve the training efficiency, prediction performance and better capture the complex relationship between features, the data needs to be standardized.

[0149] Transform the data along the time dimension to a distribution with a mean of 0 and a standard deviation of 1, eliminate the dimensional difference between different features, and improve the convergence speed and prediction performance of the model.

[0150] Mean and standard deviation calculation: calculate the mean μ and standard deviation stdev of each variable along the time dimension;

[0151] Standardize the data using the calculated mean and standard deviation, the formula is as follows:

[0152] ,

[0153]

[0154] Where, is the number of samples, is the value of the th sample, is the mean, is the standard deviation, is the standardized value;

[0155] Step five: Use the sine and cosine functions to encode the date and minute number extracted in the timestamp, and embed the periodic information. The encoded time features are spliced with the standardized running data in the feature dimension to form the input feature matrix. The spliced features are mapped to a high-dimensional space through an embedding layer, so that the model can capture complex feature relationships. Finally, the features are organized into a time series according to the time step, and input into the subsequent feature extraction module, the formula is as follows:

[0156]

[0157] Step six: The present application uses a four-branch architecture to extract time series features, and the preprocessed data is processed through four parallel branches, which are bidirectional state space branch, time convolution branch, local window transformation branch and periodic enhancement branch (see Figure 1 ). The specific implementation of each branch is as follows:

[0158] 1) Bidirectional state space branch

[0159] Bidirectional state space branch is used to capture long-term periodic dependencies, whose structure is shown in Figure 2 After embedding, the feature matrix enters the forward and reverse state space transformation 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, as follows:

[0160]

[0161] where, is the feature of the current variable, is the input feature, is the hidden state, , , , D is the training parameter;

[0162] The outputs of the forward and reverse transformations are directly added to generate the integrated features. To capture global features and long-term dependencies, multi-layer stacking is used, with residual connections and normalization layers added between each layer to enhance model stability. After processing, the features enter the feedforward network, extract features through the linear rectifier activation function, and finally compress back to the original dimension through the linear layer. This branch effectively captures the long-term trends of weekly and seasonal cycles, such as seasonal fluctuations in load variables.

[0163] 2) Time convolution branch

[0164] The time convolution branch extracts local periodic features through multi-scale dilated convolution and gating mechanism, whose structure is shown in Figure 3 The input feature is first adapted to two-dimensional convolution operation through dimension rearrangement and expansion. Multi-scale dilated convolution is used to extract features on the filter path and the gating path respectively, with a convolution kernel size of 3, as follows:

[0165]

[0166] where, represents the convolution kernel weight, represents the value of the input signal at time step , is the convolution kernel size, is the dilation factor.

[0167] The tensor after convolution output is padded to make the output dimension align with the input, avoiding the loss of boundary information, as follows:

[0168]

[0169] 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 integrated features. This branch effectively captures local features such as daily cycles.

[0170] 3) Local window attention branch

[0171] The local window transformation branch enhances short-term feature expression by limiting attention range. Its structure is shown in Figure 4 The input data is combined with a high-resolution slicing strategy to generate a high-resolution time series. The sliced data is mapped to a high-dimensional representation through an embedding layer and position embedding is added.

[0172] The data enters the multi-layer encoding layer, which calculates short-term dependencies 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:

[0173]

[0174] where, , , are the query, key, and value vectors within the window, is the dimension of the key.

[0175] Residual connections and normalization layers are used between layers to enhance stability. This branch effectively captures short-term patterns in sub-daily cycles.

[0176] 4) Periodicity enhancement branch

[0177] The periodicity enhancement branch directly processes embedded features, enhancing periodic features through linear transformation and nonlinear activation. The features are processed through two layers of linear networks and linear rectifier activation.

[0178] Step seven: integrate multi-branch features through the feature fusion module to generate integrated feature representation. The specific steps are as follows:

[0179] Concatenate the four branches along the feature dimension to generate a high-dimensional feature tensor.

[0180] Step eight: input the fused features into a single-layer Chebyshev polynomial to generate the prediction sequence. This module is implemented through the following steps:

[0181] Apply the hyperbolic tangent activation function to limit the feature values to the range [-1, 1];

[0182] Initialize the single-layer Chebyshev module, set the input dimension, output dimension, and fixed polynomial order, and define the learnable coefficient matrix. The coefficient matrix is initialized with a small-scale normal distribution to ensure training stability.

[0183] The Chebyshev polynomial basis functions of 0th order to Kth order are calculated to generate a polynomial tensor.

[0184] The Chebyshev polynomial tensor is weighted and summed with the learnable coefficient matrix through tensor operation to generate a prediction sequence.

[0185] The output sequence is reshaped to the target prediction dimension to generate a periodic prediction sequence of the oil temperature and load variables, and the original dimension is restored through a reverse normalization operation.

[0186] To verify the application effect of the present application in the real world, the hybrid model proposed in the present application is evaluated on four real data sets (ETTH1, ETTH2, ETTm1, ETTm2), and five multivariate time series prediction models in the past three years are compared. The experimental results are summarized in Table 1.

[0187] Table 1 Experimental results

[0188]

[0189] As can be seen from Table 1, compared with traditional methods or single models, the present 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 new energy access scenario. The generated high-precision periodic prediction sequence provides reliable load and oil temperature prediction for energy scheduling, supports equipment maintenance plan optimization, and improves operation and maintenance efficiency. The single-layer Chebyshev projection module reduces the parameter quantity and computational complexity through fixed order design and small-scale coefficient initialization, is suitable for multivariate time series prediction under complex operating conditions, guarantees the robustness and stability of the prediction results, and provides technical support for the stable operation of low-carbon power systems.

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 nonlinear transformation on the fused features to generate a periodic prediction sequence of oil temperature and load variables. 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.

2. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 1, 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.

3. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 1, 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.

4. The method for predicting periodic time series of power transformers based on a hybrid model according to claim 1, 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.

5. 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.

6. 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.

7. 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 6, 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