Power grid transformer bushing state prediction method based on Autoformer
By using an Autoformer-based prediction method, Fourier transform and an improved attention mechanism to filter out noise, and constructing an encoder-decoder structure, the problems of technical lag and noise interference in transformer bushing state prediction are solved, achieving higher prediction accuracy and stability.
Patent Information
- Application Number
- CN202511264882.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-05
- Publication Date
- 2025-12-19
AI Technical Summary
Existing technologies lack sufficient research on the prediction of transformer bushing conditions, especially in the application of prediction algorithms, which suffer from technological lag, insufficient extraction of complex frequency features, and significant high-frequency noise interference.
We employ an Autoformer-based prediction method, combining Fourier transform signal decomposition and an improved attention mechanism with a convolutional network to filter noise in the frequency domain, using a novel loss function to reduce the risk of overfitting, and constructing an encoder-decoder structure for prediction.
This improved the accuracy and stability of transformer bushing condition prediction, reduced noise interference, enhanced the model's ability to capture core data features, avoided the risk of overfitting, and improved the accuracy of prediction results.
Smart Images
Figure CN121167115A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of transformer bushing state prediction, in particular to a power grid transformer bushing state prediction method based on Autoformer. BACKGROUND
[0002] With the continuous development of social economy and the continuous progress of science and technology, the physical structure of modern power grid system is becoming increasingly large, and the technical complexity and operation difficulty are significantly increased. The close coupling characteristics of each link of the power grid system make any local fault have a profound impact on the stable operation of the whole system. On the other hand, the large-scale application of smart grid and renewable energy technologies has led to an exponential increase in power grid operation data, which puts higher technical requirements on the massive data processing and analysis capabilities in operation and maintenance.
[0003] Key power equipment in the power system often has the characteristics of strong burstiness and great harm. Once the equipment fails and causes operation interruption, not only will it cause significant direct economic losses, but also will seriously affect social production and normal power consumption of residents. Therefore, accurately predicting the state of the core equipment of the power grid and realizing early warning and intervention of faults are of great importance to the safety of the power grid. The transformer is one of the core equipment of the power system, and its operation reliability is directly related to the safety of the power grid. The transformer bushing is a key component that carries and conducts electric energy in the transformer, and its insulation state directly affects the normal transmission of electric energy. Statistical data shows that the faults caused by the bushing account for a significant proportion of the overall transformer accidents, usually between 9.9% and 12%, and in certain operating conditions, the proportion can be as high as 40%, which fully demonstrates the necessity and urgency of predicting the state of the transformer bushing.
[0004] However, although there are a large number of state studies on transformers and other power equipment in the prior art, there is relatively insufficient special prediction research on the state of the transformer bushing. Especially in the application of prediction algorithms, some existing research still uses traditional models such as the long short-term memory network (LSTM) proposed in 1997. At the same time, new algorithms based on attention mechanisms that have shown excellent performance in the field of time series prediction have not been fully developed in the field of transformer bushing state prediction, and there is obvious technical lag. SUMMARY
[0005] In view of the deficiencies of the prior art, the present application provides a power grid transformer bushing state prediction method based on Autoformer, which solves the problems of insufficient extraction of complex frequency characteristics of transformer bushing data and significant high-frequency noise interference in the prior art.
[0006] To achieve the above object, the present application is implemented by the following technical solutions: the present application provides a power grid transformer bushing state prediction method based on Autoformer, which comprises the following steps:
[0007] Step one: data acquisition and preprocessing
[0008] The monitoring data of the power grid transformer bushing is acquired, and the monitoring data at least includes one or more of dielectric loss, end screen current and equivalent capacitance. For the missing values in the monitoring data caused by sensor failure and the like, a cubic spline interpolation method is used for completion processing. The method divides the data into multiple small intervals, and in each interval, a cubic polynomial is fitted to generate the missing point function value, and the polynomial structure is:
[0009] y i (x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 ;
[0010] In the formula, y i (x) represents the interpolation function in the i-th interval; a i ,b i ,c i ,d i represent the coefficients to be solved of the cubic polynomial; x represents the independent variable, i.e. the time point; x i represents the starting point of the i-th interval.
[0011] The preprocessed data sequence X raw is processed by using a signal decomposition method FDR based on Fourier transform. The specific steps are as follows:
[0012] Discrete Fourier transform DFT is performed on X raw :
[0013] X f =F(X raw );
[0014] In the formula, X f is the representation of data sequence X raw in the frequency domain; F is the discrete Fourier transform operation. The energy of each frequency component is calculated:
[0015] E k =|X f (k)| 2 ;
[0016] E(k) = |X(k)|2 k E(k) = |X(k)|2 f (k) is the complex value of the component with frequency index k in the frequency domain.
[0017] According to the calculated energy size, the preset number (for example, 10) of frequency components with the highest energy are selected, and inverse discrete Fourier transform (IDFT) is performed on the selected frequency components to reconstruct the main sub-signals in the data
[0018] All reconstructed main sub-signals and residual signals S res are combined to form multi-dimensional input data X new required for subsequent modeling:
[0019]
[0020] S res is the residual signal, which is calculated as is the mth main sub-signal.
[0021] Step three: improved Autoformer model construction
[0022] An improved Autoformer prediction model is constructed, which is based on an encoder-decoder structure, and the core improvement point is:
[0023] The attention mechanism of the convolution network is integrated:
[0024] To reduce the interference of high-frequency noise in the data, the original attention mechanism of the model is improved. Its working mode is: the Q vector, K vector and V vector inside the attention mechanism are converted to the frequency domain through fast Fourier transform (FFT); in the frequency domain, a convolution network is used to process the transformed signals; after processing, correlation calculation and time delay aggregation operations are performed.
[0025] Design of new loss function:
[0026] To reduce the risk of overfitting in the model training process, a new preset loss function L DL is used:
[0027]
[0028] L DL is the calculation result of the loss function; L is the data length of the predicted sequence; y i is the predicted value of the ith point in the sequence; y new is the true value of the ith point in the sequence; α is a fine-tuning parameter; β is a learnable hyperparameter; tanh is the hyperbolic tangent function.
[0029] Step 4: Model Training and Prediction
[0030] The multidimensional input data X generated in step two new The model is divided into training and test sets according to a preset ratio (e.g., 8:2), and the improved Autoformer model is trained using the training set. The encoder-decoder structure of this model internally processes time series data through the sequence decomposition module Decomp, the aforementioned improved attention mechanism Correlation, and the feedforward network. Its core computational process includes:
[0031]
[0032] In the formula, This is the input to the l-th layer of the encoder; and These are the seasonal and trend component sub-signals obtained by the encoder in the j-th decomposition of the l-th layer, respectively. This is the input to the l-th layer of the decoder; and These are the seasonal and trend component sub-signals obtained by the decoder in the j-th decomposition of the l-th layer, respectively. This is the output of the last layer (Nth layer) of the encoder; This is the final trend component output of the decoder's layer l; This is the final seasonal component output of the decoder's layer l, and l represents the level index; j represents the decomposition count index; W l,j The learnable weight coefficients are the j-th extracted trend component of the l-th layer.
[0033] The final prediction result is generated by superimposing the seasonal and trend components of the last layer of the decoder:
[0034]
[0035] In the formula, prediction is the final sequence of prediction results; and These are the seasonal component output and trend component output of the last layer (layer L) of the decoder, respectively.
[0036] Step 5: Model Testing and Performance Evaluation
[0037] Test process and output processing:
[0038] Load the trained model file, input the preprocessed test set data, and the model generates a 3D output M. o Its dimensions are:
[0039] [batchsize, length, dim];
[0040] where batchsize is the batch number, length is the length of the batch input data, and dim is the feature dimension.
[0041] Subsequently, a dimension reduction connection operation is performed on the three-dimensional output to obtain a final one-dimensional prediction output:
[0042]
[0043] where finalprediction is the final spliced one-dimensional output sequence; M o is the model output of the oth batch; N batch is the total number of batches; flat is a data flattening dimension reduction operation, and a is the batch index.
[0044] Performance evaluation:
[0045] Through the evaluation module, the prediction performance of the model is quantitatively evaluated using at least one of the following indicators:
[0046] Mean Squared Error (MSE):
[0047]
[0048] Mean Absolute Error (MAE):
[0049]
[0050] Root Mean Squared Error (RMSE):
[0051]
[0052] Coefficient of Determination (R 2 :
[0053]
[0054] In the above evaluation indicators, L is the length of the test set data sequence; y i is the true value of the i th point; is the predicted value of the i th point; is the mean value of the true value, MSE is used to calculate the average of the squared difference between the predicted value and the actual value, MAE is used to calculate the average of the absolute difference between the predicted value and the actual value, RMSE is the square root of MSE, which is used to evaluate the general size of the prediction error, R 2 is a value between 0 and 1, indicating the percentage of variance explained by the model, SSR is the sum of squared residuals, and SST is the total sum of squares.
[0055] The application provides a power grid transformer bushing state prediction method based on an Autoformer.
[0056] 1、The application adopts a signal decomposition method FDR based on Fourier transform, decomposes and reconstructs the original transformer bushing monitoring signal containing complex frequency components into a multi-dimensional input composed of main signal components and residual signals. This enhances the model's ability to capture the core features of the data, reduces the difficulty of processing caused by non-stationary original data, and thus improves the accuracy of subsequent prediction.
[0057] 2、The application integrates a convolutional network in the attention mechanism of the Autoformer model and converts the attention calculation process to the frequency domain. Through this structure, high-frequency noise interference in the original data is filtered out at the frequency domain level, so that the model can focus more on the effective change rule in the signal, avoiding the disturbance of noise to the prediction result, and thus improving the stability and accuracy of the prediction result.
[0058] 3、The application adopts a new loss function combining mean square error and mean absolute error regularization term to replace the traditional mean square error loss function for model training. This design reduces the risk of overfitting of the model during deep training, ensures the stability and convergence effect of the model training, and thus obtains a final prediction model with better performance. BRIEF DESCRIPTION OF DRAWINGS
[0059] Figure 1 is a system flowchart of the application;
[0060] Figure 2 is a schematic diagram of the improved attention mechanism of the application;
[0061] Figure 3 is a model structure diagram of the application;
[0062] Figure 4 is an example diagram of the real transformer bushing data set (dielectric loss, end screen current, equivalent capacitance) of the application;
[0063] Figure 5 is a performance evaluation diagram of the optimal model of the specific embodiment of the application;
[0064] Figure 6 is a prediction curve result diagram of the optimal model of the specific embodiment of the application;
[0065] Figure 7 is an encoding-decoding structure diagram of the application. DETAILED DESCRIPTION
[0066] With reference to the drawings of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present application.
[0067] Please refer to the drawings of the present application Figure 1 -attached Figure 7 , the embodiment of the present application provides a power grid transformer bushing state prediction method based on Autoformer, comprising.
[0068] Step 1: Obtain data
[0069] The original data comes from 500kV transformer C-phase bushing monitoring data, the sampling interval is 20 minutes, there are 17,083 data points, and the data set contains three key parameters of dielectric loss, tail screen current and equivalent capacitance.
[0070] Step 2: Data preprocessing
[0071] For the problem of data missing caused by sensor failure, the present application uses Cubic Spline Interpolation method to complete. The steps are as follows:
[0072] Divide the data into small intervals [x i , x i+1 ], and fit a cubic polynomial in each [x i , x i+1 ].
[0073] y i (x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 Set boundary conditions to get the equation group.
[0074] Solve the coefficients (a i , b i , c i , d i ).
[0075] Based on the coefficients, generate the missing point function value S i (x miss ), and get the processed data X raw .
[0076] Step 3: Data decomposition
[0077] Reduce data complexity by FDR:
[0078] For pre-processed data X raw Perform Discrete Fourier Transform (DFT):
[0079] X f = F(X raw );
[0080] where X f is the representation of data sequence X raw in frequency domain; F is the Discrete Fourier Transform operation.
[0081] Calculate the energy E k of each frequency component (k is the frequency index):
[0082] E k = |X f (k)| 2 ;
[0083] where E k is the energy of the kth frequency component; X f (k) is the complex value of the component with frequency index k in frequency domain.
[0084] Select the top 10 sub-signals with the largest energy (X f (k max,1 ), X f (k max,2 ),..., X f (k max,10 )).
[0085] Perform IDFT transform on the selected components to reconstruct the main sub-signals
[0086]
[0087] Recombine the main sub-signals with the residual signal into multi-dimensional input,
[0088]
[0089] Step 4: Data division
[0090] Divide the real transformer bushing data that has been processed in steps 2 and 3 into training set and test set in the ratio of 8:2.
[0091] Improve the model:
[0092] Optimize the attention mechanism:
[0093] In order to reduce the high-frequency interference in the data, the attention mechanism of the original Autoformer model is improved. The Q vector, the K vector and the V vector are converted to the frequency domain, and the interference of high-frequency noise is filtered out through a convolution network. The structure is shown in the accompanying Figure 2 .
[0094] Loss function design:
[0095] A new loss function is used to replace MSE to reduce the risk of model gradient disappearance:
[0096]
[0097] In the formula, L DL is the calculation result of the loss function; L is the data length of the predicted sequence; is the predicted value of the i-th point in the sequence; y i is the true value of the i-th point in the sequence; alpha is a fine-tuning parameter; beta is a learnable hyperparameter; tanh is the hyperbolic tangent function.
[0098] Finally, the improved Autoformer prediction model is built based on the pycharm and pytorch deep learning framework. The overall structure of the model is shown in the accompanying Figure 3 .
[0099] Step 6: Train the model
[0100] Model training parameter settings:
[0101] The number of convolution output channels: 512;
[0102] The number of hidden layer units: 2048;
[0103] Optimizer: Stochastic Gradient Descent (SGD);
[0104] Learning rate: 0.001;
[0105] Training rounds: 100;
[0106] Create an encoder-decoder:
[0107] Put the input Autoformer encoder-decoder of the training set divided in step 4:
[0108]
[0109] In the formula, is the input of the l-th layer of the encoder; and are the seasonal component and trend component sub-signals obtained by the encoder in the l-th layer j-th decomposition, respectively; is the input of the l-th layer of the decoder; and are the seasonal and trend component sub-signals obtained by the decoder in the j-th decomposition of the l-th layer, respectively; is the output of the last layer (N-th layer) of the encoder; is the final trend component output of the l-th layer of the decoder; is the final seasonal component output of the l-th layer of the decoder, and l denotes the layer index; j denotes the decomposition index; W l,j is the learnable weight coefficient of the j-th extracted trend component of the l-th layer.
[0110]
[0111] where prediction is the final prediction result sequence; and are the seasonal and trend component outputs of the last layer (L-th layer) of the decoder, respectively. The above encoding-decoding process is shown in FIG. 2. Figure 7
[0112] Step 7: Test the model
[0113] Test flow:
[0114] Load the optimal model weights.
[0115] Input the preprocessed test set data.
[0116] Generate a three-dimensional output through the encoding-decoding process: [batchsize, length, dim];
[0117] where batchsize, length, and dim represent the batch size (32), the length of the batch input data (96), and the feature dimension (11), respectively.
[0118] Perform dimension reduction splicing:
[0119]
[0120] where finalprediction is the final one-dimensional output sequence after splicing; M o is the model output of the o-th batch; N batch is the total number of batches; flat is the operation of flattening and reducing the dimension of the data, and a is the batch index.
[0121] Performance evaluation:
[0122] Through the evaluation module, the prediction performance of the model is quantitatively evaluated using at least one of the following indicators:
[0123] Mean Squared Error, MSE:
[0124]
[0125] Mean Absolute Error, MAE:
[0126]
[0127] Root Mean Squared Error, RMSE:
[0128]
[0129] Coefficient of Determination, R 2 :
[0130]
[0131] In the above evaluation indicators, L is the length of the test set data sequence; y i is the true value of the i-th point; is the predicted value of the i-th point; is the mean value of the true value, MSE is used to calculate the average value of the squared difference between the predicted value and the actual value, MSE is used to calculate the average value of the absolute difference between the predicted value and the actual value, RMSE is the square root of MSE, which is used to evaluate the general size of the prediction error, R 2 is a value between 0 and 1, indicating the percentage of variance explained by the model, SSR is the sum of squared residuals, SST is the total sum of squares, and the results are shown in the attached Figure 5 .
[0132] Step 1: Data preprocessing of the original transformer bushing monitoring data provided by the power grid, the processed data is shown in the attached Figure 4 , the data contains 17083 continuous sampling points of dielectric loss, end screen current and equivalent capacitance parameters.
[0133] Step 2: Use FDR to decompose the three variables obtained above into multiple sub-signals of different frequencies, and select the top 10 sub-signals with the largest energy for each variable according to the energy size, and the residual signal constitutes a multi-dimensional signal as the data input of the model.
[0134] Step 3: The preprocessed data in step 2 (a total of 17803 samples) is randomly divided into training set and test set according to the ratio of 8:2.
[0135] Step 4: Based on pycharm and pytorch deep learning framework, an improved Autoformer prediction model is built.
[0136] Step five: the model built in step four is trained with appropriate hyperparameters and an encoder-decoder structure.
[0137] Step six: the model weights saved in step five are loaded, and the model is evaluated on the test set data divided in step three under different prediction lengths. The evaluation indicators are the four commonly used indicators in prediction tasks (MSE, MAE, RMSE, R2). The evaluation results are shown in the following table Figure 5 .
[0138] The optimal prediction model in step six is loaded, and the test set data is predicted from the transformer bushing data set that has been divided in step three. The prediction curve results are shown in the following figure Figure 6 .
Claims
1. A method for predicting the condition of power grid transformer bushings based on Autoformer, characterized in that, Includes the following steps: Acquire monitoring data of transformer bushings, the monitoring data including at least one or more of dielectric loss, end-screen current and equivalent capacitance; and perform data preprocessing on the monitoring data; The Fourier transform-based signal decomposition method (FDR) is used to decompose and reconstruct the preprocessed data into multidimensional input data. An improved Autoformer prediction model is constructed and trained using the multidimensional input data. The improved Autoformer prediction model integrates a convolutional network in the attention mechanism and adopts a pre-defined novel loss function. The improved Autoformer prediction model, after training, is used to predict the state of transformer bushings.
2. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 1, characterized in that, The data preprocessing includes: using cubic spline interpolation to fill in missing values in the monitoring data.
3. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 1, characterized in that, The signal decomposition method based on Fourier transform specifically includes: Perform a Discrete Fourier Transform (DFT) on the preprocessed data; Calculate the energy of each frequency component; Select a preset number of sub-signals with the highest energy, and perform an inverse discrete Fourier transform (IDFT) on the selected sub-signals to reconstruct the main signal components; The reconstructed main signal components are combined with the residual signal to form the multidimensional input data.
4. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 1, characterized in that, The attention mechanism that integrates a convolutional network works as follows: the Q, K, and V vectors in the attention mechanism are transformed to the frequency domain through a Fast Fourier Transform (FFT), processed by the convolutional network in the frequency domain, and then correlation calculation and time delay aggregation are performed to filter out high-frequency noise interference.
5. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 1, characterized in that, The pre-defined novel loss function is used to replace the mean squared error (MSE) loss function in order to reduce the risk of gradient vanishing during model training.
6. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 1, characterized in that, The improved Autoformer prediction model employs an encoder-decoder structure.
7. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 6, characterized in that, The encoder-decoder structure includes a sequence decomposition module, which is used to decompose the input time series into different sub-signals.
8. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 7, characterized in that, The encoder receives the sub-signals obtained by the sequence decomposition module and extracts and encodes the sub-signals using the attention mechanism that integrates a convolutional network.
9. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 7, characterized in that, The decoder also includes a sequence decomposition module and an attention mechanism that integrates a convolutional network. It decodes the encoder's output and its own input, and finally superimposes the decoded trend component and seasonal component to generate the final prediction result.
10. The method for predicting the bushing condition of a power grid transformer based on Autoformer according to claim 1, characterized in that, The casing condition prediction method further includes: after model training is completed, using a testing module, reserving test set data and selecting from preset evaluation indicators, the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used. 2 At least one of the evaluation metrics is used to evaluate the predictive performance of the trained model.