Method and device for predicting service life of transformer

Abstract

The invention provides a method and a device for predicting the service life of a transformer. The method for predicting the service life comprises the following steps: acquiring static characteristic data and dynamic characteristic data of the transformer; preprocessing the static feature data to obtain first static feature data; preprocessing the dynamic feature data to obtain first dynamic feature data; performing feature coding on the first static feature data to obtain second static feature data; performing feature coding on the first dynamic feature data to obtain second dynamic feature data; performing feature extraction on the second dynamic feature data to obtain a first predicted life; performing feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted life; and fusing the first predicted life and the second predicted life to obtain the predicted life of the transformer. According to the embodiment of the invention, the prediction result is closer to the real life, and potential faults are warned in advance.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of transformers, and in particular to a method and device for predicting the life of a transformer. Background Art

[0002] Power transformers are core equipment in power transmission and distribution systems, and their operating status directly impacts grid security. Current transformer life prediction methods include life assessment based on thermal aging models, which infer the degree of polymerization of the insulation paper from the hotspot temperature of the windings. However, these methods ignore the coupled effects of multiple factors, such as mechanical stress and partial discharge. Other methods use Dissolved Gas Analysis (DGA) combined with the IEC (three characteristic gases) three-ratio method to determine fault type based on dissolved gas content. However, these methods suffer from data lag and blind spots in threshold determination. Statistical models based on the Weibull distribution rely on historical fault databases and are difficult to adapt to the diverse designs of transformers. Summary of the Invention

[0003] The technical problem to be solved by the embodiments of the present invention is to provide a method and device for predicting the life of a transformer, which can make the prediction result closer to the actual life and provide early warning of potential faults.

[0004] To solve the above technical problems, the technical solutions of the embodiments of the present invention are as follows: A transformer life prediction method, comprising: Obtaining static characteristic data and dynamic characteristic data of the transformer; Preprocessing the static feature data to obtain first static feature data; Preprocessing the dynamic feature data to obtain first dynamic feature data; Performing feature encoding on the first static feature data to obtain second static feature data; Performing feature encoding on the first dynamic feature data to obtain second dynamic feature data; performing feature extraction on the second dynamic feature data to obtain a first predicted lifespan; performing feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifespan; The first predicted life and the second predicted life are combined to obtain the predicted life of the transformer.

[0005] Optionally, the static characteristic data includes: rated capacity, ambient temperature, ambient humidity, voltage level and / or insulation material; The dynamic characteristic data include: winding leakage current, winding dielectric loss factor, insulation resistance absorption ratio and / or polarization index.

[0006] Optionally, preprocessing the static feature data to obtain first static feature data includes: Outliers are eliminated from the static feature data and the dynamic feature data to obtain first static feature data.

[0007] Optionally, preprocessing the dynamic feature data to obtain first dynamic feature data includes: Outlier elimination, missing value completion, and data enhancement are performed on the dynamic feature data to obtain first dynamic feature data.

[0008] Optionally, performing feature encoding on the first dynamic feature data to obtain second dynamic feature data includes: Applying a sliding window to the first dynamic feature data to obtain a statistical sequence that changes over time; Converting the time-varying statistical sequence into a high-frequency energy ratio; All high-frequency energy proportions are summarized to obtain the second dynamic characteristic data.

[0009] Optionally, performing feature extraction on the second dynamic feature data to obtain a first predicted lifespan includes: according to , determine the first predicted life span; in, RUL phy is the first predicted lifespan, L 0 is the initial lifespan, u for t 0- t The instantaneous time interval in a moment, D is the temperature sensitivity coefficient, x ( u ) is the real-time second dynamic feature data, x 0 is the reference second dynamic feature data.

[0010] Optionally, performing feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifespan includes: determining a first weight according to the second dynamic characteristic data, wherein the first weight is a thermal conduction weight, an electromagnetic coupling weight, or a mechanical coupling weight; determining an attention score according to the second dynamic feature data; Determine a second weight based on the first weight and the attention score; determining a state vector of each component according to the second dynamic feature data and the second weight; determining a comprehensive representation of global features based on the second dynamic feature data and the state vectors of each component; A second predicted lifespan is determined based on the comprehensive representation of the global characteristics and the second dynamic characteristic data.

[0011] Optionally, determining a second predicted lifespan based on the comprehensive representation of the global characteristics and the second dynamic characteristic data includes: Determine a hidden state vector based on the comprehensive representation of the global features and the second dynamic feature data; determining an embedding vector of a degradation stage according to the second dynamic feature data; concatenating the hidden state vector with the embedding vector in the degradation phase to obtain a concatenated vector; determining a lifespan residual according to the splicing vector; A second predicted lifespan is determined based on the lifespan residual and the current lifespan.

[0012] Optionally, fusing the first predicted life and the second predicted life to obtain the predicted life of the transformer includes: Determining an allocation weight based on the hidden state vector and the second static feature data; The predicted life of the transformer is determined according to the first predicted life, the second predicted life and the allocated weight.

[0013] An embodiment of the present invention further provides a transformer life prediction device, comprising: An acquisition module, used to acquire static characteristic data and dynamic characteristic data of the transformer; a processing module configured to preprocess the static feature data to obtain first static feature data; preprocess the dynamic feature data to obtain first dynamic feature data; feature encode the first static feature data to obtain second static feature data; feature encode the first dynamic feature data to obtain second dynamic feature data; feature extract the second dynamic feature data to obtain a first predicted life; feature extract the second static feature data and the second dynamic feature data to obtain a second predicted life; and fuse the first predicted life and the second predicted life to obtain a predicted life of the transformer.

[0014] The above solution of the embodiment of the present invention has at least the following beneficial effects: The above-mentioned scheme of the embodiment of the present invention simultaneously incorporates static characteristic data (inherent properties such as rated capacity, ambient temperature and humidity, voltage level and insulation material) and dynamic characteristic data (real-time status parameters such as winding leakage current, dielectric loss factor and insulation resistance absorption ratio), thereby achieving a comprehensive characterization of the design properties and operating status of the transformer, avoiding the one-sidedness caused by a single factor or data type.

[0015] Through the process of preprocessing-feature encoding, the original data is cleaned, standardized and converted, which solves the compatibility problem of different types of features and reduces the interference of noise data and outliers on the prediction results.

[0016] The first predicted lifespan is obtained by extracting dynamic features alone, focusing on the impact of real-time operating status on lifespan; the second predicted lifespan is obtained by combining static and dynamic features, reflecting the coupling effect of inherent properties and real-time status; finally, the advantages of both are combined through a fusion strategy, which not only retains the timeliness of dynamic data but also incorporates the differentiation basis of static data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a transformer life prediction method provided by an embodiment of the present invention.

[0018] Figure 2 Schematic diagram of the remaining life prediction percentage and error of Example 1 provided by an embodiment of the present invention; Figure 3 It is a module schematic diagram of the transformer life prediction device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0020] like Figure 1 As shown, an embodiment of the present invention provides a transformer life prediction method, comprising: Step 11, obtaining static characteristic data and dynamic characteristic data of the transformer; Step 12: preprocessing the static feature data to obtain first static feature data; Step 13, preprocessing the dynamic feature data to obtain first dynamic feature data; Step 14, performing feature encoding on the first static feature data to obtain second static feature data; Step 15, performing feature encoding on the first dynamic feature data to obtain second dynamic feature data; Step 16, performing feature extraction on the second dynamic feature data to obtain a first predicted lifespan; Step 17, performing feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifespan; Step 18: Merge the first predicted life and the second predicted life to obtain the predicted life of the transformer.

[0021] Specifically, the static characteristic data includes: rated capacity, ambient temperature, ambient humidity, voltage level and / or insulation material; The dynamic characteristic data include: winding leakage current, winding dielectric loss factor, insulation resistance absorption ratio and / or polarization index.

[0022] In this embodiment, static characteristic data (inherent properties such as rated capacity, ambient temperature and humidity, voltage level and insulation material) and dynamic characteristic data (real-time status parameters such as winding leakage current, dielectric loss factor and insulation resistance absorption ratio) are included at the same time, achieving a comprehensive characterization of the design properties and operating status of the transformer, avoiding the one-sidedness caused by a single factor or data type.

[0023] Through the process of preprocessing-feature encoding, the original data is cleaned, standardized and converted, which solves the compatibility problem of different types of features and reduces the interference of noise data and outliers on the prediction results.

[0024] The first predicted lifespan is obtained by extracting dynamic features alone, focusing on the impact of real-time operating status on lifespan; the second predicted lifespan is obtained by combining static and dynamic features, reflecting the coupling effect of inherent properties and real-time status; finally, the advantages of both are combined through a fusion strategy, which not only retains the timeliness of dynamic data but also incorporates the differentiation basis of static data.

[0025] In an optional embodiment of the present invention, in step 12, preprocessing the static feature data to obtain first static feature data includes: Step 121 : Eliminate abnormal values ​​from the static feature data and the dynamic feature data to obtain first static feature data.

[0026] In step 13, the dynamic feature data is preprocessed to obtain first dynamic feature data, including: Step 131 : performing outlier elimination, missing value filling, and data enhancement on the dynamic feature data to obtain first dynamic feature data.

[0027] Specifically, eliminating outliers from the static feature data and the dynamic feature data may include: according to , Determine anomaly scores for each static feature data and dynamic feature data; Among them, the s ( x ) are the anomaly scores of static feature data and dynamic feature data, E ( h ( x )) is the average path length of static feature data and dynamic feature data in the forest, h (x ) is the path length of static feature data and dynamic feature data in a single tree, x =1, 2, ..., n , n is the total sample size of static feature data and dynamic feature data, c ( n ) is the expected path length of static feature data and dynamic feature data; in, , c is Euler's constant; Determine whether the anomaly score is less than a set threshold; if so, exclude the corresponding static feature data or dynamic feature data.

[0028] Filling missing values ​​in the dynamic feature data may include: according to , Determine the missing value completion results of each dynamic feature data; in, The missing value completion results for each dynamic feature data, i =1, 2, ..., m , m is the total amount of missing value completion results, j =1, 2, ..., K , K is the total number of neighbor samples, KNN ( i ) is a sample with missing values i The set of neighbor samples of w ij For missing value samples i The nearest neighbor sample j The weight of x j For the nearest neighbor sample j The known eigenvalues ​​of in, , d ( x i , x j ) is a sample with missing values i With neighbor samples j distance, exp is an exponential function, o is the characteristic variance of the set of neighbor samples.

[0029] Performing data enhancement on the dynamic feature data (transformer insulation degradation data and fault data) may include: according to , Determining first enhanced data of first dynamic characteristic data in the dynamic characteristic data (insulation paper aging related data, such as polymerization degree and remaining life); in, L ( t )for t Remaining life (or insulation residual performance) at the moment, L 0 is the initial lifespan, a is the first material constant, β is the second material constant, i 0 is the rated reference temperature of the winding, i ( u ) is the real-time temperature of the winding, u 0- t The instantaneous time interval in the moment; according to the first dynamic characteristic data (initial life L 0), by adjusting i 0, get L ( t ) changes over time, i 0 and L ( t ) is converted into the first enhanced data (such as monthly remaining life and aging rate) to supplement the lack of real samples; according to , Determining second enhanced data of second dynamic characteristic data in the dynamic characteristic data (low-probability fault data such as a sudden increase in dielectric loss factor and abnormal dissolved gas content in oil); in, V ( D , G ) is a generator G and the discriminator D The loss function is For the generator G Minimize loss function, discriminator D Maximize the loss function, x : Pdata is the data selected from the real fault sample library, D ( x ) is the discriminator D judge x is the probability of true data, logD ( x ) is to convert the probability into loss, E x：Pdata [ logD ( x )] is to average the losses of all real data; z : Pz For the generator G Random noise in the input, G ( z ) is a generator G Use noise z Generated fake fault samples, D ( G ( z )) is the discriminator D judge G ( z ) is the probability of detection of real data, log (1- D ( G ( z ))) To turn the probability of detection into loss, E z：Pz [ log (1- D ( G ( z )))] is to average the losses of all generated fake data; The data obtained after outlier exclusion, missing value filling and data enhancement are aggregated to obtain the first static feature data and the first dynamic feature data.

[0030] In this embodiment, the isolation forest algorithm is used to identify abnormal samples (such as a sudden increase in leakage current), eliminate interference data, reduce the impact of extreme values ​​on model training, and improve data reliability.

[0031] The weighted KNN interpolation method (combining the exponential weight formula of sample distance and feature variance) is used for dynamic feature data to fill in missing values ​​through weighted calculation of neighboring samples.

[0032] Based on the physical aging model formula, by adjusting the winding reference temperature, we generate data on the change in remaining life over time (such as monthly aging rate), supplementing the lack of real insulation paper aging samples. Using a generative adversarial network (GAN), through the game between the generator and the discriminator, we generate rare fault samples such as sudden increases in dielectric loss factor and abnormal gas in oil.

[0033] In an optional embodiment of the present invention, in step 14, feature encoding is performed on the first static feature data to obtain second static feature data. Specifically, discrete classification data (such as insulating materials) in the first static feature data is converted into a 0-1 vector. For example, for three materials (A, B, C), A is encoded as [1, 0, 0], B is encoded as [0, 1, 0], and C is encoded as [0, 0, 1]. according to , Standardizing continuous numerical data (such as rated capacity) in the first static characteristic data; in, y is continuous numerical data, m is the mean of continuous numerical data, d is the standard deviation of continuous numerical data, is the standardized continuous numerical data; The 0-1 vector and the normalized continuous numerical data are aggregated to obtain second static feature data.

[0034] In step 15, feature encoding is performed on the first dynamic feature data to obtain second dynamic feature data, including: Step 151: Apply a sliding window to the first dynamic feature data to obtain a time-varying statistical sequence. Specifically, apply a sliding window (taking 24 consecutive hours of current data) to the first dynamic feature data (e.g., leakage current) to obtain the mean and variance of the first dynamic feature data within each sliding window. The means and variances of all first dynamic feature data are aggregated to obtain a time-varying statistical sequence. Step 152: transform the time-varying statistical sequence into a high-frequency energy ratio; specifically, according to , transform the time-varying statistical sequence into the proportion of high-frequency energy; in, W f ( f ) is the wavelet coefficient, representing the statistical sequence that changes over time at different frequencies f The weight on For frequencies exceeding the cutoff frequency f 0 (corresponding to the high-frequency signal of partial discharge), which reflects the high-frequency energy of sudden faults. is the sum of squares of wavelet coefficients of all frequencies, reflecting the total energy, E h is the proportion of high-frequency energy, reflecting the degree of significance of high-frequency fault signals; Step 153: Summarize all high-frequency energy proportions to obtain second dynamic characteristic data.

[0035] In this embodiment, one-hot encoding is used for discrete data (such as insulation materials) to eliminate the influence of the implicit order of numerical values; continuous data (such as rated capacity) is standardized and dimensioned to avoid imbalance in model weights due to differences in feature scales and improve the consistency of feature input.

[0036] Sliding window statistical extraction (mean, variance): Captures long-term trends in dynamic parameters (such as leakage current), smoothes sensor noise, and highlights slow degradation patterns.

[0037] Wavelet transform generates high-frequency energy proportions, identifies early high-frequency signals of sudden faults (such as partial discharge), enhances sensitivity to transient anomalies, and provides early warning of potential faults.

[0038] In an optional embodiment of the present invention, in step 16, feature extraction is performed on the second dynamic feature data to obtain a first predicted lifespan, including: Step 161, according to , determine the first predicted life span; in, RUL phy is the first predicted lifespan, L 0 is the initial lifespan, u for t 0- t The instantaneous time interval in a moment, D is the temperature sensitivity coefficient, x ( u ) is the real-time second dynamic characteristic data (winding real-time temperature), x 0 is the reference second dynamic characteristic data (winding rated reference temperature).

[0039] In this embodiment, the difference between the winding temperature and the reference temperature is used to quantify the impact of thermal aging, ensuring that the prediction results conform to the physical laws of insulation material aging and providing a highly interpretable white box benchmark for subsequent fusion.

[0040] In an optional embodiment of the present invention, in step 17, feature extraction is performed on the second static feature data and the second dynamic feature data to obtain a second predicted lifespan, including: Step 171: determining a first weight according to the second dynamic characteristic data; specifically, the first weight may be a thermal conduction weight, an electromagnetic coupling weight, or a mechanical coupling weight; Step 1711, according to , determine the influence coefficient of temperature on heat conduction; in, w θ is the influence coefficient of temperature on heat conduction, E a is the activation energy, with a typical value of 8.617×10 -5 eV / K, k is the Boltzmann constant, i is the winding temperature, i 0 is the reference temperature; Step 1712, according to determining a first weight; in, w pq is the first weight, which is the heat conduction weight, bis the thermal conductivity coefficient, DT p is the first dynamic characteristic of the first component, DT q is a second dynamic characteristic of the second component; Step 1713, according to determining a first weight; Wherein, the first weight is the electromagnetic coupling weight, m is the angular frequency of the alternating current; Step 1714, according to determining a first weight; Wherein, the first weight is the mechanical coupling weight, t is the mechanical coupling coefficient; Step 172: determine the attention score based on the second dynamic feature data; specifically, , determine the attention score; in, e pq is the attention score, a is the first learnable parameter, W 1 is the second learnable parameter, W 2 is the third learnable parameter; Step 173: determining a second weight according to the first weight and the attention score; Specifically, according to determining a second weight; in, A pq is the second weight, g =1, 2, 3..., R , R is the total number of components in the second dynamic feature data; Step 174: determining a state vector of each component based on the second dynamic feature data and the second weight; Specifically, according to Determine the state vector of each component; in, B p is the state vector of each component, s is the activation function, W v is the learnable value matrix; Step 175: determining a comprehensive representation of the global feature based on the second dynamic feature data and the state vectors of each component; Specifically, according to Get the global characteristics of each component; in, DT p （t）is the state characteristic of each component at time t, GRU is a gated recurrent unit, DT p （t-1） is the state characteristics of each component at time t-1; according to ,determine the comprehensive representation of global features; in, f graph is a comprehensive representation of global features. C p is the component importance weight; Step 176 : Determine a second predicted lifespan based on the comprehensive representation of the global characteristics and the second dynamic characteristic data.

[0041] In this embodiment, the multi-physics coupling relationship between components (such as thermal conduction between windings and insulating oil, and electromagnetic coupling between windings and iron core) is quantified, and physical constraints are used to ensure the comprehensiveness of component interaction modeling, avoiding the problem of purely data-driven models ignoring physical mechanisms.

[0042] The gated attention mechanism dynamically adjusts the component interaction weights (such as increasing the impact on insulating oil when the winding temperature rises), allowing the model to adaptively capture key coupling relationships under real-time working conditions and improve its sensitivity to complex dynamic changes.

[0043] GRU is used to capture the temporal dependency of component degradation, weightedly aggregate global features, and fuse multi-component state information to provide comprehensive device-level feature input for the black box model.

[0044] LSTM captures dynamic time series trends, and HMM explicitly identifies insulation degradation stages (normal / degraded / faulty), improving the ability to capture nonlinear degradation patterns, supplementing the impact of complex operating conditions not covered by the physical model, and enhancing prediction robustness.

[0045] In an optional embodiment of the present invention, in step 176, determining a second predicted lifespan based on the comprehensive representation of the global characteristics and the second dynamic characteristic data includes: Step 1761, determining a hidden state vector based on the comprehensive representation of the global feature and the second dynamic feature data; Specifically, according to , determine the hidden state vector, in, s t is the hidden state vector at the current moment, LSTM is the long short-term memory network operator, s t-1 is the hidden state vector at the previous moment, f graph is a comprehensive representation of global features. DT t is the second dynamic characteristic data (leakage current and dielectric loss factor); Step 1762: Determine an embedding vector of the degradation stage based on the second dynamic feature data; Specifically, according to , determine the degradation stage, in, q t is the degradation stage at the current moment (including normal stage, degraded stage and fault stage), q t-1 is the degradation stage of the previous moment, A is the state transition matrix of the second dynamic feature data, inferring the current degradation stage; The degradation stage at the current moment q t Embedding vector, get the embedding vector of the degradation stage ; Step 1763: concatenate the hidden state vector with the embedding vector in the degradation phase to obtain a concatenated vector. Specifically, according to , determine the lifetime residual, in, c t is the concatenation vector, Concat is the splicing operation function; Step 1764, determining the life residual according to the splicing vector; Specifically, according to , determine the lifetime residual, in, △RUL DL is the life residual, FC is a fully connected layer; Step 1765, determining a second predicted lifespan based on the lifespan residual and the current lifespan; Specifically, according to , determine the second predicted lifespan, in, RUL DL is the second predicted lifespan, RUL 1 is the current lifespan (which can be the first predicted lifespan or set based on historical data).

[0046] In this embodiment, step 1761 uses an LSTM model to process global features (component interaction information) and dynamic time series data (such as leakage current and dielectric loss factor) to capture the nonlinear evolution of device state over time. The LSTM memory mechanism effectively preserves long-term degradation trends (such as the slow aging of insulation materials) while responding to short-term fluctuations (such as sudden temperature rises caused by load changes). This provides a continuous and dynamic feature representation for subsequent degradation stage identification and lifespan prediction, avoiding the problem of traditional time series models' inability to capture long-term dependencies.

[0047] Step 1762 explicitly divides insulation degradation into three stages (normal, degraded, and faulty) using the HMM model. The state transition matrix quantifies the transition probabilities between stages (e.g., the trigger conditions for transitioning from normal to degraded). This stage information is converted into an embedded vector, allowing the abstract degradation state to be directly processed by the deep learning model, enhancing sensitivity to fault precursors (e.g., subtle characteristics of the initial stages of partial discharge).

[0048] Step 1763 combines the continuous state features captured by the LSTM with the discrete stage information identified by the HMM, achieving complementary alignment between micro-series trends and macro-stage positioning. For example, when the hidden state vector indicates a slow temperature increase (a long-term trend) while the embedded vector indicates a degradation stage (a macro-state), the concatenated data can more accurately reflect the complex characteristics of accelerated aging in the degradation stage, providing more comprehensive input for subsequent residual prediction and reducing information loss from a single feature dimension.

[0049] Step 1764 performs a nonlinear transformation on the concatenated vectors using a fully connected layer, outputting the lifetime residual, which quantifies the deviation between the current state and the baseline lifetime. This residual prediction mechanism adaptively corrects the theoretical values ​​of the physical model (such as the additional lifetime loss due to mechanical stress), compensating for the purely physical model's neglect of complex coupling factors (such as the synergistic effects of vibration and high temperature), and improving its adaptability to actual operating conditions.

[0050] The residuals are combined with the current lifespan to generate the final black-box model prediction. This baseline + correction model ensures that the prediction results are anchored to physical laws (avoiding outrageous values ​​driven purely by data) while also adapting to the differentiated degradation characteristics of individual devices through residual correction, balancing prediction stability with personalized needs.

[0051] In an optional embodiment of the present invention, in step 18, the first predicted life and the second predicted life are combined to obtain the predicted life of the transformer, including: Step 181, determining an allocation weight according to the hidden state vector and the second static feature data; Specifically, according to , determine the allocation weight; in, l To assign weights, sis the activation function, w T is the transpose of the weight vector, s t is the hidden state vector at the current moment, m is the second static feature data; Step 182: determining a predicted life of the transformer according to the first predicted life, the second predicted life, and the assigned weight; Specifically, according to , determine the predicted life of the transformer, in, RUL To predict life expectancy.

[0052] In this embodiment, the weights are dynamically adjusted based on static features (such as insulating materials) and dynamic states (hidden vectors). The weights of the white box model are enhanced in scenarios dominated by physical laws (such as normal operating conditions), and the weights of the black box model are enhanced in complex nonlinear scenarios (such as multi-factor coupling), thereby achieving a balance between accuracy and interpretability.

[0053] Combining the stability of physical models with the flexibility of deep learning, the limitations of a single model are reduced, and prediction accuracy is improved (for example, the error after fusion in this embodiment is lower than that of traditional LSTM), providing a more reliable basis for operation and maintenance decisions.

[0054] In an optional embodiment of the present invention, the life prediction method further includes: Step 19, determining a prediction error based on the predicted lifespan and the actual lifespan according to the method of the present invention; Specifically, according to , determine the prediction error; in, MAE is the prediction error, e =1,2,3.. N , N is the number of predictions, RUL ε is the current predicted lifespan, RUL ε true is the current real lifespan; Step 20, comparing the prediction error of the method of the present invention with the prediction error of the traditional LSTM (Long Short-Term Memory Network) method.

[0055] In this embodiment, by MAE The prediction deviation was quantified and compared with the traditional LSTM to verify the model's effectiveness. The proposed method's prediction results are closer to the actual lifespan and provide early warning of potential failures, avoiding safety risks caused by delayed predictions. This provides a scientific basis for power companies to plan equipment replacements in advance, ensuring the safe operation of the power grid.

[0056] Example 1 For a 220kV oil-immersed power transformer with a rated capacity of 100MVA, insulation material of Class A paper, and an operating environment temperature of 25°C and humidity of 60%, Example 1 provides a transformer life prediction method, including: Step 21: Obtain static and dynamic characteristic data of the transformer. The static characteristic data are: rated capacity 100 MVA, voltage level 220 kV, insulation material grade A paper, ambient temperature 25°C, and ambient humidity 60%. The dynamic characteristic data of the monitoring data over the past 30 days are: winding leakage current: 0.5-1.2 mA; winding dielectric loss factor (tan δ): 0.3%-0.8%; insulation resistance absorption ratio: 1.2-1.8; and polarization index: 2.0-3.5. Step 22: Perform outlier elimination, missing value filling, and data enhancement on the static feature data and dynamic feature data to obtain first static feature data and first dynamic feature data. For example, if the leakage current suddenly increases to 2 mA, the abnormal sample is eliminated. The missing dielectric loss factor data is filled using the KNN interpolation method with a 5-day window. Based on the IEEE C57.91 standard, the insulation degradation process when the winding temperature increases from 80°C to 120°C is simulated to generate 1000 sets of simulation data. 100 sets of partial discharge fault samples are generated using a generative adversarial network. Step 23: Feature encode the first static feature data and the first dynamic feature data to obtain second static feature data and second dynamic feature data. The first static feature data is encoded, for example, the insulation material type is grade A paper → One-Hot encoding is [1, 0, 0], where the material is grade A / B / C respectively; the voltage level is 220 kV → One-Hot encoding is [0, 1, 0], where the transformer voltage level is 110 kV / 220 kV / 500 kV; the ambient temperature (25°C) is normalized to 25 − 20 / 5 = 1.0; the first dynamic feature data is enhanced, for example, the leakage current 24-hour window captures the trend of parameter changes over time, and the mean is 0.8 mA, the variance is 0.02; the dielectric loss factor trend slope is 0.005% / day; Step 24: Use the first model to extract the second static characteristic data and the second dynamic characteristic data to obtain a first predicted lifespan. Assuming that the transformer design lifespan is L 0 is 20 years, D =0.05, current temperature x ( u ) = 25℃, the predicted remaining life is 18 years.

[0057] Step 25, using the second model to perform feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifespan; according to , determine the influence coefficient of temperature on heat conduction; where, i =25℃, i 0=20℃; Enter component characteristics in the physical constraint layer: Winding (N1): [Temperature = 25°C, Leakage Current = 0.8 mA, Tanδ = 0.5%]; Iron Core (N2): [Vibration Frequency = 100 Hz, Magnetic Flux Density = 1.5 T]; Insulating Oil (N3): [pH = 5.5, Moisture Content = 20 ppm]; Generate the interaction weight matrix:

[0058] where w 13 =0.6 is the thermal conductivity weight of the winding → insulating oil; w 12 =0.8 is the electromagnetic coupling weight of winding ↔ core.

[0059] In the gated attention calculation mechanism section:

[0060]

[0061]

[0062] Compute dynamic state evolution, such as capturing degradation trends due to increasing temperature: B 1 = [0.9, 0.6, 0.3], B 2 = [0.7, 0.5, 0.2], B 3 = [0.6, 0.4, 0.1], the weighted average global feature is:

[0063] LSTM predicted state hidden state vector s t , HMM identifies the degradation stage and outputs RUL DL =16 years; Step 26: Merge the first predicted life and the second predicted life to obtain a predicted life of the transformer; , ; Step 27, conduct experimental verification and compare the predicted results of this method with the real life span and the traditional LSTM prediction method. The results are as follows Figure 2The results show that the lifespan prediction method constructed by this invention is closer to the actual lifespan of power transformers. Furthermore, these predicted lifespans are generally earlier than the actual lifespans, allowing power companies to replace power transformers in advance, thereby preventing potential serious failures. In contrast, the lifespans predicted using the LSTM method deviate significantly from the actual lifespans, and most of the predicted lifespans are later than the actual lifespans. This can affect the normal operation of the power system and, in severe cases, even threaten the safety of maintenance personnel, resulting in irreversible losses.

[0064] The present invention uses the isolation forest algorithm to identify abnormal samples (such as a sudden increase in leakage current), eliminate interfering data, reduce the impact of extreme values ​​on model training, and improve data reliability.

[0065] The weighted KNN interpolation method (combining the exponential weight formula of sample distance and feature variance) is used for dynamic feature data to fill in missing values ​​through weighted calculation of neighboring samples.

[0066] Based on the physical aging model formula, by adjusting the winding reference temperature, we generate data on the change in remaining life over time (such as monthly aging rate), supplementing the lack of real insulation paper aging samples. Using a generative adversarial network (GAN), through the game between the generator and the discriminator, we generate rare fault samples such as sudden increases in dielectric loss factor and abnormal gas in oil.

[0067] One-hot encoding is used for discrete features (such as insulation materials A / B / C) to eliminate the influence of implicit numerical order.

[0068] Standardize continuous features (such as rated capacity) and unify the dimensions (such as ambient temperature is standardized to 1.0) to avoid imbalance in model weights due to scale differences and ensure fair input of static features (such as voltage level and insulation material).

[0069] Sliding window statistics (mean, variance) are extracted to capture long-term degradation trends of parameters such as leakage current (e.g., a mean of 0.8 mA and a variance of 0.02 reflect slow degradation) and smooth noise.

[0070] Wavelet transform generates high-frequency energy proportions: This identifies early high-frequency signals of sudden faults such as partial discharge, enhances sensitivity to transient anomalies, and provides early warning of potential faults (e.g., an increase in the proportion of high-frequency components of partial discharge).

[0071] Based on the physical formula of thermal aging, the impact of the difference between the real-time winding temperature and the reference temperature on the lifespan is quantified. The results anchor the physical laws of insulation material aging, provide an explainable white-box benchmark for prediction, and avoid outrageous values ​​driven by pure data.

[0072] Multi-physics coupling weights (thermal conduction, electromagnetic coupling, and mechanical coupling) are introduced to quantify the interaction between components (e.g., the thermal conduction weight of winding → insulating oil is 0.6), solving the problem that traditional models ignore the synergistic effects of multiple factors.

[0073] Combining LSTM to capture time series trends and HMM to divide degradation stages (normal / degraded / faulty), life residuals are dynamically corrected to adapt to the differentiated degradation characteristics of individual transformers (such as accelerated aging of new materials).

[0074] Weights are adaptively assigned based on static features (such as insulating materials) and dynamic states (hidden vectors), enhancing the weight of physical models under normal operating conditions and the weight of data-driven models in complex coupling scenarios.

[0075] The fusion results are anchored to physical laws (avoiding deviations) and adapt to actual working conditions (such as additional loss due to mechanical stress). They are closer to the actual lifespan than traditional LSTM predictions, and provide sufficient advance warning time, reserving a buffer period for operation and maintenance decisions.

[0076] pass MAE Quantization error,Compared with the traditional LSTM method, the prediction results of the present invention are closer to the actual lifespan, and the predicted values ​​are generally earlier than the actual lifespan, allowing power companies to plan replacements in advance and prevent serious failures.

[0077] Traditional LSTM ignores physical constraints and component interactions, resulting in large prediction deviations and even lagging behind the actual lifespan, which may lead to failure risks.

[0078] like Figure 3 As shown, an embodiment of the present invention further provides a transformer life prediction device 30, comprising: An acquisition module 31 is used to acquire static characteristic data and dynamic characteristic data of the transformer; The processing module 32 is used to preprocess the static feature data to obtain first static feature data; preprocess the dynamic feature data to obtain first dynamic feature data; feature encode the first static feature data to obtain second static feature data; feature encode the first dynamic feature data to obtain second dynamic feature data; feature extract the second dynamic feature data to obtain a first predicted life; feature extract the second static feature data and the second dynamic feature data to obtain a second predicted life; and fuse the first predicted life and the second predicted life to obtain a predicted life of the transformer.

[0079] Optionally, the static characteristic data includes: rated capacity, ambient temperature, ambient humidity, voltage level and / or insulation material; The dynamic characteristic data include: winding leakage current, winding dielectric loss factor, insulation resistance absorption ratio and / or polarization index.

[0080] Optionally, preprocessing the static feature data to obtain first static feature data includes: Outliers are eliminated from the static feature data and the dynamic feature data to obtain first static feature data.

[0081] Optionally, preprocessing the dynamic feature data to obtain first dynamic feature data includes: Outlier elimination, missing value completion, and data enhancement are performed on the dynamic feature data to obtain first dynamic feature data.

[0082] Optionally, performing feature encoding on the first dynamic feature data to obtain second dynamic feature data includes: Applying a sliding window to the first dynamic feature data to obtain a statistical sequence that changes over time; Converting the time-varying statistical sequence into a high-frequency energy ratio; All high-frequency energy proportions are summarized to obtain the second dynamic characteristic data.

[0083] Optionally, performing feature extraction on the second dynamic feature data to obtain a first predicted lifespan includes: according to , determine the first predicted life span; in, RUL phy is the first predicted lifespan, L 0 is the initial lifespan, u for t 0- t The instantaneous time interval in a moment, D is the temperature sensitivity coefficient, x ( u ) is the real-time second dynamic feature data, x 0 is the reference second dynamic feature data.

[0084] Optionally, performing feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifespan includes: determining a first weight according to the second dynamic characteristic data, wherein the first weight is a thermal conduction weight, an electromagnetic coupling weight, or a mechanical coupling weight; determining an attention score according to the second dynamic feature data; Determine a second weight based on the first weight and the attention score; determining a state vector of each component according to the second dynamic feature data and the second weight; determining a comprehensive representation of global features based on the second dynamic feature data and the state vectors of each component; A second predicted lifespan is determined based on the comprehensive representation of the global characteristics and the second dynamic characteristic data.

[0085] Optionally, determining a second predicted lifespan based on the comprehensive representation of the global characteristics and the second dynamic characteristic data includes: Determine a hidden state vector based on the comprehensive representation of the global features and the second dynamic feature data; determining an embedding vector of a degradation stage according to the second dynamic feature data; concatenating the hidden state vector with the embedding vector in the degradation phase to obtain a concatenated vector; determining a lifespan residual according to the splicing vector; A second predicted lifespan is determined based on the lifespan residual and the current lifespan.

[0086] Optionally, fusing the first predicted life and the second predicted life to obtain the predicted life of the transformer includes: Determining an allocation weight based on the hidden state vector and the second static feature data; The predicted life of the transformer is determined according to the first predicted life, the second predicted life and the allocated weight.

[0087] It should be noted that this device is a device corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0088] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A transformer life prediction method, characterized in that: include: Obtaining static characteristic data and dynamic characteristic data of the transformer; Preprocessing the static feature data to obtain first static feature data; Preprocessing the dynamic feature data to obtain first dynamic feature data; Performing feature encoding on the first static feature data to obtain second static feature data; Performing feature encoding on the first dynamic feature data to obtain second dynamic feature data; performing feature extraction on the second dynamic feature data to obtain a first predicted lifespan; performing feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifespan; The first predicted life and the second predicted life are combined to obtain the predicted life of the transformer.

2. The transformer life prediction method according to claim 1, characterized in that: The static characteristic data includes: rated capacity, ambient temperature, ambient humidity, voltage level and / or insulation material; The dynamic characteristic data include: winding leakage current, winding dielectric loss factor, insulation resistance absorption ratio and / or polarization index.

3. The transformer life prediction method according to claim 1, characterized in that: Preprocessing the static feature data to obtain first static feature data includes: Outliers are eliminated from the static feature data and the dynamic feature data to obtain first static feature data.

4. The transformer life prediction method according to claim 1, characterized in that: Preprocessing the dynamic feature data to obtain first dynamic feature data includes: Outlier elimination, missing value completion, and data enhancement are performed on the dynamic feature data to obtain first dynamic feature data.

5. The transformer life prediction method according to claim 1, characterized in that: Performing feature encoding on the first dynamic feature data to obtain second dynamic feature data includes: Applying a sliding window to the first dynamic feature data to obtain a statistical sequence that changes over time; Converting the time-varying statistical sequence into a high-frequency energy ratio; All high-frequency energy proportions are summarized to obtain the second dynamic characteristic data.

6. The transformer life prediction method according to claim 1, characterized in that: Extracting features from the second dynamic feature data to obtain a first predicted lifespan includes: according to , determine the first predicted life span; in, RUL phy is the first predicted lifespan, L 0 is the initial lifespan, u for t 0- t The instantaneous time interval in a moment, D is the temperature sensitivity coefficient, x ( u ) is the real-time second dynamic feature data, x 0 is the reference second dynamic feature data.

7. The transformer life prediction method according to claim 6, characterized in that: Extracting features from the second static feature data and the second dynamic feature data to obtain a second predicted lifespan includes: determining a first weight according to the second dynamic characteristic data, wherein the first weight is a thermal conduction weight, an electromagnetic coupling weight, or a mechanical coupling weight; determining an attention score according to the second dynamic feature data; Determine a second weight based on the first weight and the attention score; determining a state vector of each component according to the second dynamic feature data and the second weight; determining a comprehensive representation of global features based on the second dynamic feature data and the state vectors of each component; A second predicted lifespan is determined based on the comprehensive representation of the global characteristics and the second dynamic characteristic data.

8. The transformer life prediction method according to claim 7, characterized in that: Determining a second predicted lifespan based on the comprehensive representation of the global characteristics and the second dynamic characteristic data includes: Determine a hidden state vector based on the comprehensive representation of the global features and the second dynamic feature data; determining an embedding vector of a degradation stage according to the second dynamic feature data; concatenating the hidden state vector with the embedding vector in the degradation phase to obtain a concatenated vector; determining a lifespan residual according to the splicing vector; A second predicted lifespan is determined based on the lifespan residual and the current lifespan.

9. The transformer life prediction method according to claim 8, characterized in that: The first predicted life and the second predicted life are combined to obtain the predicted life of the transformer, including: Determining an allocation weight based on the hidden state vector and the second static feature data; The predicted life of the transformer is determined according to the first predicted life, the second predicted life and the allocated weight.

10. A transformer life prediction device, characterized in that: include: An acquisition module, used to acquire static characteristic data and dynamic characteristic data of the transformer; a processing module configured to preprocess the static feature data to obtain first static feature data; preprocess the dynamic feature data to obtain first dynamic feature data; feature encode the first static feature data to obtain second static feature data; feature encode the first dynamic feature data to obtain second dynamic feature data; feature extract the second dynamic feature data to obtain a first predicted life; feature extract the second static feature data and the second dynamic feature data to obtain a second predicted life; and fuse the first predicted life and the second predicted life to obtain a predicted life of the transformer.

Images