A method and apparatus for predicting the life of a transformer

Abstract

The application provides a transformer life prediction method and device, the life prediction method comprises: obtaining static characteristic data and dynamic characteristic data of a transformer; preprocessing the static characteristic data to obtain first static characteristic data; preprocessing the dynamic characteristic data to obtain first dynamic characteristic data; performing feature coding on the first static characteristic data to obtain second static characteristic data; performing feature coding on the first dynamic characteristic data to obtain second dynamic characteristic data; performing feature extraction on the second dynamic characteristic data to obtain a first predicted life; performing feature extraction on the second static characteristic data and the second dynamic characteristic data to obtain a second predicted life; and fusing the first predicted life and the second predicted life to obtain a predicted life of the transformer. The embodiment of the application can make the prediction result closer to the real life and give an early warning of potential failure.

Description

Technical Field

[0001] The present invention relates to the field of transformer technology, and in particular to a method and apparatus for predicting the lifespan of a transformer. Background Technology

[0002] As a core piece of equipment in power transmission and distribution systems, the operating status of power transformers directly affects grid security. Current methods for predicting transformer lifespan include life assessment based on thermal aging models, which calculates the degree of polymerization of insulating paper based on winding hot spot temperatures, but neglects the coupled effects of multiple factors such as mechanical stress and partial discharge. Other methods use oil chromatography (DGA) combined with the IEC (three characteristic gases) three-ratio method to determine fault type based on dissolved gas content, but suffer from data lag and threshold determination blind spots. Still other methods rely on statistical models based on the Weibull distribution, depending on historical fault databases, making it difficult to adapt to the differentiated 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 apparatus for predicting the life of a transformer, which can make the prediction results closer to the actual life and provide early warning of potential faults.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] A method for predicting the lifespan of a transformer, comprising:

[0006] Obtain static and dynamic characteristic data of the transformer;

[0007] The static feature data is preprocessed to obtain the first static feature data;

[0008] The dynamic feature data is preprocessed to obtain the first dynamic feature data;

[0009] The first static feature data is encoded to obtain the second static feature data;

[0010] The first dynamic feature data is encoded to obtain the second dynamic feature data;

[0011] Feature extraction is performed on the second dynamic feature data to obtain the first predicted lifetime;

[0012] Feature extraction is performed on the second static feature data and the second dynamic feature data to obtain the second predicted lifetime;

[0013] The predicted lifetime of the transformer is obtained by fusing the first and second predicted lifetimes.

[0014] Optionally, the static characteristic data includes: rated capacity, ambient temperature, ambient humidity, voltage level and / or insulation material;

[0015] The dynamic characteristic data include: winding leakage current, winding dielectric loss factor, insulation resistance absorption ratio, and / or polarization index.

[0016] Optionally, the static feature data is preprocessed to obtain first static feature data, including:

[0017] Outlier exclusion is performed on the static and dynamic feature data to obtain the first static feature data.

[0018] Optionally, the dynamic feature data is preprocessed to obtain first dynamic feature data, including:

[0019] The dynamic feature data is subjected to outlier removal, missing value completion, and data augmentation to obtain the first dynamic feature data.

[0020] Optionally, feature encoding is performed on the first dynamic feature data to obtain second dynamic feature data, including:

[0021] A sliding window is applied to the first dynamic feature data to obtain a sequence of statistics that change over time;

[0022] The time-varying statistical sequence is transformed into a high-frequency energy percentage.

[0023] By summing up the proportions of all high-frequency energies, we obtain the second dynamic characteristic data.

[0024] Optionally, feature extraction is performed on the second dynamic feature data to obtain a first predicted lifetime, including:

[0025] according to Determine the first predicted lifetime;

[0026] in, RUL phy For the first predicted lifespan, L 0 represents the initial lifespan. u for t 0- t The instantaneous time interval in a moment, D For temperature sensitivity coefficient, x ( u This represents real-time second dynamic feature data. x 0 represents the second dynamic feature data for reference.

[0027] Optionally, feature extraction is performed on the second static feature data and the second dynamic feature data to obtain the second predicted lifetime, including:

[0028] A first weight is determined based on the second dynamic feature data, wherein the first weight is a heat conduction weight, an electromagnetic coupling weight, or a mechanical coupling weight.

[0029] The attention score is determined based on the second dynamic feature data;

[0030] The second weight is determined based on the first weight and the attention score;

[0031] Based on the second dynamic feature data and the second weight, determine the state vector of each component;

[0032] Based on the second dynamic feature data and the state vectors of each component, a comprehensive representation of the global features is determined;

[0033] The second predicted lifetime is determined based on the comprehensive characterization of the global features and the second dynamic feature data.

[0034] Optionally, the second predicted lifetime is determined based on the comprehensive representation of the global features and the second dynamic feature data, including:

[0035] Based on the comprehensive representation of the global features and the second dynamic feature data, the hidden state vector is determined;

[0036] Based on the second dynamic feature data, determine the embedding vector for the degradation stage;

[0037] The hidden state vector is concatenated with the embedding vector of the degradation stage to obtain the concatenated vector;

[0038] The lifetime residual is determined based on the splicing vector.

[0039] The second predicted lifetime is determined based on the lifetime residual and the current lifetime.

[0040] Optionally, the first predicted lifetime and the second predicted lifetime are fused to obtain the predicted lifetime of the transformer, including:

[0041] The weights are determined based on the hidden state vector and the second static feature data.

[0042] The predicted life of the transformer is determined based on the first predicted life, the second predicted life, and the assigned weights.

[0043] Embodiments of the present invention also provide a transformer life prediction device, comprising:

[0044] The acquisition module is used to acquire the static and dynamic characteristic data of the transformer.

[0045] The processing module is configured to preprocess the static feature data to obtain first static feature data; preprocess the dynamic feature data to obtain first dynamic feature data; perform feature encoding on the first static feature data to obtain second static feature data; perform feature encoding on the first dynamic feature data to obtain second dynamic feature data; perform feature extraction on the second dynamic feature data to obtain a first predicted lifetime; perform feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifetime; and fuse the first predicted lifetime and the second predicted lifetime to obtain the predicted lifetime of the transformer.

[0046] The above-described solutions of the embodiments of the present invention have at least the following beneficial effects:

[0047] The above-described solution of this invention incorporates both 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 transformer's design attributes and operating status and avoiding the one-sidedness caused by a single factor or data type.

[0048] By employing a preprocessing-feature encoding process, the raw data is cleaned, standardized, and transformed, thus resolving compatibility issues between different types of features and reducing the interference of noisy data and outliers on the prediction results.

[0049] The first predicted lifetime is obtained by extracting dynamic features alone, focusing on the impact of real-time operating status on lifetime; the second predicted lifetime is obtained by combining static and dynamic features, reflecting the coupling effect of inherent attributes and real-time status; finally, the advantages of both are combined through a fusion strategy, which not only preserves the timeliness of dynamic data, but also incorporates the differentiated foundation of static data. Attached Figure Description

[0050] Figure 1 This is a flowchart of a transformer life prediction method provided in an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the remaining lifetime prediction percentage and error in Example 1 provided by the embodiments of the present invention;

[0052] Figure 3 This is a schematic diagram of the transformer life prediction device provided in an embodiment of the present invention. Detailed Implementation

[0053] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0054] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting the lifespan of a transformer, comprising:

[0055] Step 11: Obtain the static and dynamic characteristic data of the transformer;

[0056] Step 12: Preprocess the static feature data to obtain the first static feature data;

[0057] Step 13: Preprocess the dynamic feature data to obtain the first dynamic feature data;

[0058] Step 14: Perform feature encoding on the first static feature data to obtain the second static feature data;

[0059] Step 15: Perform feature encoding on the first dynamic feature data to obtain the second dynamic feature data;

[0060] Step 16: Extract features from the second dynamic feature data to obtain the first predicted lifetime;

[0061] Step 17: Extract features from the second static feature data and the second dynamic feature data to obtain the second predicted lifetime;

[0062] Step 18: Combine the first predicted lifetime and the second predicted lifetime to obtain the predicted lifetime of the transformer.

[0063] Specifically, the static characteristic data includes: rated capacity, ambient temperature, ambient humidity, voltage level and / or insulation material;

[0064] The dynamic characteristic data include: winding leakage current, winding dielectric loss factor, insulation resistance absorption ratio, and / or polarization index.

[0065] In this embodiment, both 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 incorporated, realizing a comprehensive characterization of the transformer's design attributes and operating status, and avoiding the one-sidedness caused by a single factor or data type.

[0066] By employing a preprocessing-feature encoding process, the raw data is cleaned, standardized, and transformed, thus resolving compatibility issues between different types of features and reducing the interference of noisy data and outliers on the prediction results.

[0067] The first predicted lifetime is obtained by extracting dynamic features alone, focusing on the impact of real-time operating status on lifetime; the second predicted lifetime is obtained by combining static and dynamic features, reflecting the coupling effect of inherent attributes and real-time status; finally, the advantages of both are combined through a fusion strategy, which not only preserves the timeliness of dynamic data, but also incorporates the differentiated foundation of static data.

[0068] In an optional embodiment of the present invention, step 12 involves preprocessing the static feature data to obtain first static feature data, including:

[0069] Step 121: Outlier exclusion is performed on the static feature data and dynamic feature data to obtain the first static feature data.

[0070] In step 13, the dynamic feature data is preprocessed to obtain the first dynamic feature data, including:

[0071] Step 131: Perform outlier removal, missing value completion, and data augmentation on the dynamic feature data to obtain the first dynamic feature data.

[0072] Specifically, outlier removal from the static and dynamic feature data may include:

[0073] according to ,

[0074] Determine the anomaly scores for each static and dynamic feature data;

[0075] Among them, the s ( x () represents the anomaly scores for static and dynamic feature data. E ( h ( x The average path length of static and dynamic feature data in the forest is represented by the given path length. h ( x ) represents the path length of static and dynamic feature data within a single tree. x =1, 2, ..., n , n This represents the total sample size of both static and dynamic feature data. c ( n () represents the expected path length of static and dynamic feature data;

[0076] in, ,

[0077] c Here is Euler's constant;

[0078] Determine whether the abnormal score is less than a set threshold. If it is less than a threshold, exclude the corresponding static or dynamic feature data.

[0079] Imposing missing values ​​on the dynamic feature data may include:

[0080] according to ,

[0081] Determine the missing value completion results for each dynamic feature data;

[0082] in, Complete the missing values ​​for each dynamic feature data. i =1, 2, ..., m , m The total amount of results for filling in missing values. j =1, 2, ..., K , K The total number of nearest neighbor samples, KNN ( i () represents samples with missing values i The set of nearest neighbor samples, w ij For missing value samples i nearest neighbor samples j The weight, x j Nearest neighbor samples j The known eigenvalues;

[0083] in, ,

[0084] d ( x i , x j () represents samples with missing values i Nearest neighbor samples j distance, exp It is an exponential function. o Let V be the characteristic variance of the set of nearest neighbor samples.

[0085] Data augmentation of the dynamic characteristic data (transformer insulation degradation data, fault data) may include:

[0086] according to ,

[0087] Determine the first enhanced data of the first dynamic feature data in the dynamic feature data (insulating paper aging related data, such as degree of polymerization and remaining life).

[0088] in, L ( t )for t The remaining lifetime (or remaining insulation performance) at any given time. L 0 represents the initial lifespan. a The first material constant, β This is the second material constant. i 0 represents the rated reference temperature of the winding. i ( u (This refers to the real-time temperature of the winding.) u 0- t Instantaneous time intervals within a given moment; based on the first dynamic characteristic data (initial lifetime) L 0), by adjusting i 0, get L ( t Data on changes over time will i 0 and L ( t The dynamic relationship between the data is transformed into first augmentation data (such as monthly remaining lifespan and aging rate) to supplement the lack of real samples.

[0089] according to ,

[0090] The second enhanced data of the second dynamic feature data in the dynamic feature data is determined (low-probability fault data such as sudden increase in medium loss factor and abnormal content of dissolved gas in oil).

[0091] in, V ( D , G ) is a generator G and discriminator D loss function,

[0092] For generator G Minimize loss function, discriminator D Maximize the loss function,

[0093] x : Pdata The data was selected from a real fault sample database. D ( x () is the discriminator D judge x It is the probability of real data. logD ( x To convert probability into loss, E x：Pdata [ logD ( x [This is to average the loss of all real data;]

[0094] z : Pz For generator G Input random noise, G ( z ) is a generator G noise z The generated fake fault samples, D ( G ( z () is the discriminator D judge G ( z () represents the probability of detection based on real data. log (1- D ( G ( z To convert the probability of detection into a loss, E z：Pz [ log (1- D ( G ( z To average the loss of all generated fake data;

[0095] The data obtained after outlier removal, missing value completion, and data augmentation are summarized to obtain the first static feature data and the first dynamic feature data.

[0096] In this embodiment, the isolated forest algorithm is used to identify abnormal samples (such as sudden increases in leakage current), remove interfering data, reduce the impact of extreme values ​​on model training, and improve data reliability.

[0097] For dynamic feature data, a weighted KNN interpolation method (combining the exponential weight formula of sample distance and feature variance) is used to fill in missing values ​​by weighted calculation of nearest neighbor samples.

[0098] Based on the physical aging model formula, by adjusting the winding reference temperature, data on the change of remaining life over time (such as monthly aging rate) is generated to supplement the deficiencies of real insulation paper aging samples. A generative adversarial network (GAN) is employed to generate rare fault samples such as sudden increases in dielectric loss factor and abnormal gas in oil through a game between the generator and discriminator.

[0099] In an optional embodiment of the present invention, in step 14, the first static feature data is feature encoded to obtain the second static feature data. Specifically, the 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].

[0100] according to ,

[0101] The continuous numerical data (such as rated capacity) in the first static characteristic data are standardized.

[0102] in, y For continuous numerical data, m The mean of continuous numerical data. d The standard deviation of continuous numerical data. It is standardized continuous numerical data;

[0103] The 0-1 vector and the standardized continuous numerical data are summarized to obtain the second static feature data.

[0104] In step 15, the first dynamic feature data is feature-encoded to obtain the second dynamic feature data, including:

[0105] Step 151: Apply a sliding window to the first dynamic feature data to obtain a statistical sequence that changes over time; specifically, apply a sliding window (take current data for 24 consecutive hours) to the first dynamic feature data (such as leakage current) to obtain the mean and variance of the first dynamic feature data in each sliding window, and summarize the mean and variance of all the first dynamic feature data to obtain a statistical sequence that changes over time.

[0106] Step 152: Transform the time-varying statistical sequence into a high-frequency energy proportion; specifically, according to... This transforms the time-varying statistical sequence into a high-frequency energy percentage.

[0107] in, W f ( f () represents wavelet coefficients, which represent a time-varying sequence of statistics at different frequencies. f The amount on,

[0108] For frequencies exceeding the cutoff frequency f The sum of squares of wavelet coefficients of 0 (corresponding to the high-frequency signal of partial discharge) reflects the high-frequency energy of sudden faults.

[0109] The sum of squares of the wavelet coefficients at all frequencies reflects the total energy.

[0110] E h The percentage of high-frequency energy reflects the degree of prominence of high-frequency fault signals;

[0111] Step 153: Summarize the proportions of all high-frequency energies to obtain the second dynamic characteristic data.

[0112] In this embodiment, discrete data (such as insulating materials) is encoded using One-Hot encoding to eliminate the influence of implicit numerical order; continuous data (such as rated capacity) is standardized to unify the units of measurement, avoid model weight imbalance due to differences in feature scale, and improve the consistency of feature input.

[0113] Sliding window extraction of statistics (mean, variance): captures long-term trends of dynamic parameters (such as leakage current), smooths sensor noise, and highlights slow degradation patterns.

[0114] 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.

[0115] In an optional embodiment of the present invention, step 16, performing feature extraction on the second dynamic feature data to obtain a first predicted lifetime, includes:

[0116] Step 161, according to Determine the first predicted lifetime;

[0117] in, RUL phy For the first predicted lifespan, L 0 represents the initial lifespan. u for t 0- t The instantaneous time interval in a moment, D For temperature sensitivity coefficient, x ( u The second dynamic characteristic data (real-time winding temperature) is real-time. x 0 is the reference second dynamic characteristic data (winding rated reference temperature).

[0118] In this embodiment, the difference between the winding temperature and the reference temperature is used to quantify the effect of thermal aging, ensuring that the prediction results conform to the physical laws of insulation material aging, and providing a highly interpretable white-box reference for subsequent integration.

[0119] In an optional embodiment of the present invention, step 17, performing feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifetime, includes:

[0120] Step 171: Determine the first weight based on the second dynamic feature data; specifically, the first weight may be a thermal conduction weight, an electromagnetic coupling weight, or a mechanical coupling weight.

[0121] Step 1711, according to Determine the coefficient of influence of temperature on heat conduction;

[0122] in, w θThe coefficient representing the effect of temperature on heat conduction. E a The activation energy is typically taken as 8.617 × 10⁻⁶. -5 eV / K, k Boltzmann's constant, i For winding temperature, i 0 is the reference temperature;

[0123] Step 1712, according to Determine the first weight;

[0124] in, w pq The first weight is the heat conduction weight. b The thermal conductivity coefficient, DT p The first dynamic feature of the first component DT q This is the second dynamic feature of the second component;

[0125] Step 1713, according to Determine the first weight;

[0126] Wherein, the first weight is the electromagnetic coupling weight. m The angular frequency of alternating current;

[0127] Step 1714, according to Determine the first weight;

[0128] Wherein, the first weight is the mechanical coupling weight. t The mechanical coupling coefficient;

[0129] Step 172: Determine the attention score based on the second dynamic feature data; specifically, based on... Determine the attention score;

[0130] in, e pq For attention score, a As the first learnable parameter, W 1 is the second learnable parameter. W 2 is the third learnable parameter;

[0131] Step 173: Determine the second weight based on the first weight and the attention score;

[0132] Specifically, according to Determine the second weight;

[0133] in, A pq As the second weight, g =1, 2, 3...,R , R The total number of components in the second dynamic feature data;

[0134] Step 174: Determine the state vector of each component based on the second dynamic feature data and the second weight;

[0135] Specifically, according to Determine the state vector of each component;

[0136] in, B p These are the state vectors of each component. s For activation function, W v It is a learnable value matrix;

[0137] Step 175: Determine the comprehensive representation of the global features based on the second dynamic feature data and the state vectors of each component;

[0138] Specifically, according to Obtain the global features of each component;

[0139] in, DT p （t） Let t represent the state characteristics of each component at time t. GRU For gated loop unit, DT p （t-1） The state characteristics of each component at time t-1;

[0140] according to To determine a comprehensive representation of global features;

[0141] in, f graph It is a comprehensive representation of global features. C p Assign importance weights to components;

[0142] Step 176: Determine the second predicted lifetime based on the comprehensive characterization of the global features and the second dynamic feature data.

[0143] In this embodiment, the multi-physics coupling relationship between components is quantified (such as thermal conduction between the winding and the insulating oil, and electromagnetic coupling between the winding and the iron core). Physical constraints are used to ensure the comprehensiveness of component interaction modeling and avoid the problem of pure data-driven models ignoring physical mechanisms.

[0144] The gating attention mechanism dynamically adjusts the interaction weights of components (such as enhancing the influence on insulating oil when the winding temperature rises), enabling the model to adaptively capture key coupling relationships under real-time operating conditions and improve its sensitivity to complex dynamic changes.

[0145] By capturing the temporal dependencies of component degradation through GRU, weighted aggregation of global features, and fusion of multi-component state information, comprehensive device-level feature input is provided for the black-box model.

[0146] LSTM captures dynamic time-series trends, HMM explicitly identifies insulation degradation stages (normal / deterioration / fault), improves the ability to capture nonlinear degradation modes, supplements the influence of complex operating conditions not covered by the physical model, and enhances prediction robustness.

[0147] In an optional embodiment of the present invention, step 176, determining the second predicted lifetime based on the comprehensive characterization of the global features and the second dynamic feature data, includes:

[0148] Step 1761: Determine the hidden state vector based on the comprehensive representation of the global features and the second dynamic feature data;

[0149] Specifically, according to Determine the hidden state vector.

[0150] in, s t Let be the hidden state vector at the current time step. LSTM For Long Short-Term Memory network operators, s t-1 Let be the hidden state vector from the previous time step. f graph It is a comprehensive representation of global features. DT t The second dynamic characteristic data (leakage current and dielectric loss factor);

[0151] Step 1762: Determine the embedding vector for the degradation stage based on the second dynamic feature data;

[0152] Specifically, according to Determine the stage of degradation.

[0153] in, q t This refers to the current degradation stage (including the normal stage, deterioration stage, and failure stage). q t-1 This is the degenerative phase of the previous moment. A The state transition matrix of the second dynamic feature data is used to infer the current degradation stage;

[0154] The current degradation stage q t Embedding vectors, obtaining the embedding vectors of the degradation stage. ;

[0155] Step 1763: Concatenate the hidden state vector with the embedding vector of the degradation stage to obtain the concatenated vector;

[0156] Specifically, according to Determine the lifetime residual.

[0157] in, c t To concatenate vectors, Concat This is the function for concatenation operations;

[0158] Step 1764: Determine the lifetime residual based on the splicing vector;

[0159] Specifically, according to Determine the lifetime residual.

[0160] in, △RUL DL For life residual, FC It is a fully connected layer;

[0161] Step 1765: Determine the second predicted lifetime based on the lifetime residual and the current lifetime;

[0162] Specifically, according to Determine the second predicted lifetime.

[0163] in, RUL DL For the second predicted lifespan, RUL 1 represents the current lifespan (which can be the first predicted lifespan or set based on historical data).

[0164] 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 the device state over time. The memory mechanism of LSTM can effectively retain long-term degradation trends (such as slow aging of insulating materials) while responding to short-term fluctuations (such as sudden temperature rises caused by load changes), providing continuous and dynamic feature representations for subsequent degradation stage identification and lifetime prediction, avoiding the problem of traditional time-series models failing to capture long-term dependencies.

[0165] Step 1762 explicitly divides insulation degradation into three stages (normal, deterioration, and fault) using an HMM model, and quantifies the transition probabilities between stages (such as the triggering conditions from normal to deterioration) using a state transition matrix. The stage information is transformed into an embedding vector, allowing the abstract degradation state to be directly processed by the deep learning model, enhancing the sensitivity to fault precursors (such as subtle features in the early stages of partial discharge).

[0166] Step 1763 fuses the continuous state features captured by LSTM with the discrete stage information identified by HMM to achieve complementarity between micro-time series trends and macro-stage localization. For example, when the hidden state vector shows a slow temperature increase (long-term trend) and the embedded vector is labeled as a degradation stage (macro-state), the concatenated data can more accurately reflect the composite features of accelerated aging in the degradation stage, providing a more comprehensive input for subsequent residual prediction and reducing information loss from a single feature dimension.

[0167] Step 1764 performs a nonlinear transformation on the splicing vector through a fully connected layer to output the lifetime residual, quantifying the deviation between the current state and the baseline lifetime. The residual prediction mechanism can adaptively correct the theoretical values ​​of the physical model (such as the additional lifetime loss due to mechanical stress), compensate for the neglect of complex coupling factors (such as the combined effect of vibration and high temperature) by the pure physical model, and improve its adaptability to actual working conditions.

[0168] By combining the residuals with the current lifetime, the final black-box model prediction is generated. This baseline + correction approach 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, thus balancing the stability of the prediction with personalized needs.

[0169] In an optional embodiment of the present invention, step 18, fusing the first predicted lifetime and the second predicted lifetime to obtain the predicted lifetime of the transformer, includes:

[0170] Step 181: Determine the allocation weights based on the hidden state vector and the second static feature data;

[0171] Specifically, according to Determine the allocation weights;

[0172] in, l To assign weights, s For activation function, w T This is the transpose of the weight vector. s t Let be the hidden state vector at the current time step. m This is the second static feature data;

[0173] Step 182: Determine the predicted lifetime of the transformer based on the first predicted lifetime, the second predicted lifetime, and the assigned weights;

[0174] Specifically, according to To determine the predicted lifespan of the transformer,

[0175] in, RUL To predict lifespan.

[0176] 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), while 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.

[0177] By combining the stability of physical models with the flexibility of deep learning, the limitations of a single model are reduced, and the prediction accuracy is improved (e.g., the error after fusion is lower than that of traditional LSTM in the example), providing a more reliable basis for operation and maintenance decisions.

[0178] In an optional embodiment of the present invention, the lifetime prediction method further includes:

[0179] Step 19: Determine the prediction error by comparing the predicted lifetime with the actual lifetime according to the method of the present invention;

[0180] Specifically, according to Determine the prediction error;

[0181] in, MAE For prediction error, e =1, 2, 3... N , N To predict the number of times, RUL ε For current lifespan prediction, RUL ε true This represents the current actual lifespan.

[0182] Step 20: Compare the prediction error of the method of the present invention with the prediction error of the traditional LSTM (Long Short-Term Memory) method.

[0183] In this embodiment, through MAE The prediction bias is quantified, and the effectiveness of the model is verified by comparison with the traditional LSTM method. The method of this invention provides prediction results that are closer to the actual lifespan and gives early warning of potential faults, avoiding safety risks caused by prediction lag. It provides a scientific basis for power companies to plan equipment replacement in advance and ensures the safe operation of the power grid.

[0184] Example 1

[0185] For a 220kV oil-immersed power transformer with a rated capacity of 100MVA, insulation material of Class A paper, and operating ambient temperature of 25℃ and humidity of 60%, Example 1 provides a method for predicting the lifespan of a transformer, including:

[0186] Step 21: Obtain the static and dynamic characteristic data of the transformer; the static characteristic data are: rated capacity 100MVA, voltage level 220kV, insulation material Class A paper, ambient temperature 25℃, ambient humidity 60%; the dynamic characteristic data of the monitoring data over the past 30 days are: winding leakage current: 0.5~1.2mA; winding dielectric loss factor (tanδ): 0.3%~0.8%; insulation resistance absorption ratio: 1.2~1.8; polarization index: 2.0~3.5;

[0187] Step 22: Perform outlier removal, missing value completion, and data augmentation on the static and dynamic feature data to obtain the first static feature data and the first dynamic feature data; if the leakage current suddenly increases to 2mA, the abnormal sample is removed; using a 5-day window, the missing dielectric loss factor data is filled using the KNN interpolation method; based on the IEEE C57.91 standard, the insulation degradation process when the winding temperature rises from 80℃ to 120℃ is simulated to generate 1000 sets of simulation data; 100 sets of partial discharge fault samples are generated through an adversarial generative network.

[0188] Step 23: Perform feature encoding on the first static feature data and the first dynamic feature data to obtain the second static feature data and the second dynamic feature data; encode the first static feature data, such as insulation material type A-grade paper → One-Hot encoding as [1,0,0], where the material is A / B / C grade; voltage level 220kV → One-Hot encoding as [0,1,0], where the transformer voltage level is 110kV / 220kV / 500kV; ambient temperature (25℃) → standardized as 25−20 / 5=1.0; enhance the first dynamic feature data, such as leakage current 24-hour window, capture the trend of parameter change over time to obtain mean 0.8mA, variance 0.02; dielectric loss factor trend slope is 0.005% / day;

[0189] Step 24: Use the first model to extract features from the second static feature data and the second dynamic feature data to obtain the first predicted lifetime, assuming the transformer's design lifetime. L 0 represents 20 years. D =0.05, current temperature x ( u =25℃, predicted remaining lifespan is 18 years.

[0190] Step 25: Use the second model to extract features from the second static feature data and the second dynamic feature data to obtain the second predicted lifetime;

[0191] according to Determine the coefficient of influence of temperature on heat conduction; among which, i =25℃, i0 = 20℃;

[0192] Input component characteristics in the physical constraint layer: Winding (N1): [Temperature = 25℃, Leakage Current = 0.8mA, tanδ = 0.5%]; Core (N2): [Vibration Frequency = 100Hz, Magnetic Flux Density = 1.5T]; Insulating Oil (N3): [pH = 5.5, Moisture Content = 20ppm];

[0193] Generate the interaction weight matrix:

[0194]

[0195] Where w 13 =0.6 is the heat conduction weight from winding to insulating oil; w 12 =0.8 is the electromagnetic coupling weight of winding to core.

[0196] In the gating attention calculation mechanism:

[0197]

[0198]

[0199]

[0200] Calculate dynamic state evolution, such as capturing the degradation trend caused by rising 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:

[0201]

[0202] LSTM predicts the hidden state vector s t HMM identifies the degradation stage and outputs... RUL DL =16 years;

[0203] Step 26: Combine the first predicted lifetime and the second predicted lifetime to obtain the predicted lifetime of the transformer. , ;

[0204] Step 27: Conduct experimental verification by comparing the predicted results of the proposed method with the actual lifetime and the prediction results of the traditional LSTM method. The results are as follows: Figure 2As shown in the figure. The results indicate that the predicted lifespan method constructed in this invention is closer to the actual lifespan of power transformers, and these predicted lifespans are generally earlier than the actual lifespan, allowing power companies to replace power transformers in advance, thereby preventing potential serious failures. In contrast, the predicted lifespans using the LSTM method deviate significantly from the actual lifespan, and most predicted lifespans are later than the actual lifespan. This may affect the normal operation of the power system, and in severe cases, may even threaten the safety of maintenance personnel, causing irreversible losses.

[0205] This invention uses the isolated forest algorithm to identify abnormal samples (such as sudden increases in leakage current), eliminates interfering data, reduces the impact of extreme values ​​on model training, and improves data reliability.

[0206] For dynamic feature data, a weighted KNN interpolation method (combining the exponential weight formula of sample distance and feature variance) is used to fill in missing values ​​by weighted calculation of nearest neighbor samples.

[0207] Based on the physical aging model formula, by adjusting the winding reference temperature, data on the change of remaining life over time (such as monthly aging rate) is generated to supplement the deficiencies of real insulation paper aging samples. A generative adversarial network (GAN) is employed to generate rare fault samples such as sudden increases in dielectric loss factor and abnormal gas in oil through a game between the generator and discriminator.

[0208] One-Hot encoding is used for discrete features (such as insulating materials A / B / C) to eliminate the influence of the implicit order of values.

[0209] Standardize continuous features (such as rated capacity) and unify the dimensions (such as standardizing ambient temperature 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).

[0210] Sliding window extraction of statistics (mean, variance): captures the long-term degradation trend of parameters such as leakage current (e.g., 0.8mA mean and 0.02 variance reflect slow deterioration) and smooths noise.

[0211] Wavelet transform generates high-frequency energy proportion: it can identify early high-frequency signals of sudden faults such as partial discharge, enhance the sensitivity to transient anomalies, and provide early warning of potential faults (such as the increased proportion of high-frequency components in partial discharge).

[0212] 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, providing an interpretable white-box benchmark for prediction and avoiding outrageous values ​​driven by pure data.

[0213] By introducing multi-physics coupling weights (thermal conduction, electromagnetic coupling, and mechanical coupling), the interaction effects of components are quantified (e.g., the thermal conduction weight of winding to insulating oil is 0.6), thus solving the problem that traditional models ignore the synergistic effects of multiple factors.

[0214] By combining LSTM to capture timing trends with HMM to divide degradation stages (normal / deterioration / failure), lifetime residuals are dynamically corrected to adapt to the differentiated degradation characteristics of individual transformers (such as accelerated aging of new materials).

[0215] Based on static features (such as insulating materials) and dynamic states (hidden vectors), weights are adaptively assigned to enhance the weights of the physical model under normal operating conditions and enhance the weights of the data-driven model in complex coupled scenarios.

[0216] The fusion results are both anchored to physical laws (avoiding deviations) and adapted to actual working conditions (such as additional losses due to mechanical stress). They are closer to the actual lifespan than traditional LSTM predictions and provide sufficient early warning time, reserving a buffer period for operation and maintenance decisions.

[0217] pass MAE In terms of quantification error, compared with the traditional LSTM method, the prediction results of this 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.

[0218] Traditional LSTMs, by ignoring physical constraints and component interactions, have large prediction biases that may even lag behind the actual lifespan, potentially leading to failure risks.

[0219] like Figure 3 As shown, embodiments of the present invention also provide a transformer life prediction device 30, comprising:

[0220] The acquisition module 31 is used to acquire the static and dynamic characteristic data of the transformer;

[0221] 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; perform feature encoding on the first static feature data to obtain second static feature data; perform feature encoding on the first dynamic feature data to obtain second dynamic feature data; perform feature extraction on the second dynamic feature data to obtain a first predicted lifetime; perform feature extraction on the second static feature data and the second dynamic feature data to obtain a second predicted lifetime; and fuse the first predicted lifetime and the second predicted lifetime to obtain the predicted lifetime of the transformer.

[0222] Optionally, the static characteristic data includes: rated capacity, ambient temperature, ambient humidity, voltage level and / or insulation material;

[0223] The dynamic characteristic data include: winding leakage current, winding dielectric loss factor, insulation resistance absorption ratio, and / or polarization index.

[0224] Optionally, the static feature data is preprocessed to obtain first static feature data, including:

[0225] Outlier exclusion is performed on the static and dynamic feature data to obtain the first static feature data.

[0226] Optionally, the dynamic feature data is preprocessed to obtain first dynamic feature data, including:

[0227] The dynamic feature data is subjected to outlier removal, missing value completion, and data augmentation to obtain the first dynamic feature data.

[0228] Optionally, feature encoding is performed on the first dynamic feature data to obtain second dynamic feature data, including:

[0229] A sliding window is applied to the first dynamic feature data to obtain a sequence of statistics that change over time;

[0230] The time-varying statistical sequence is transformed into a high-frequency energy percentage.

[0231] By summing up the proportions of all high-frequency energies, we obtain the second dynamic characteristic data.

[0232] Optionally, feature extraction is performed on the second dynamic feature data to obtain a first predicted lifetime, including:

[0233] according to Determine the first predicted lifetime;

[0234] in, RUL phy For the first predicted lifespan, L 0 represents the initial lifespan. u for t 0- t The instantaneous time interval in a moment, D For temperature sensitivity coefficient, x ( u This represents real-time second dynamic feature data. x 0 represents the second dynamic feature data for reference.

[0235] Optionally, feature extraction is performed on the second static feature data and the second dynamic feature data to obtain the second predicted lifetime, including:

[0236] A first weight is determined based on the second dynamic feature data, wherein the first weight is a heat conduction weight, an electromagnetic coupling weight, or a mechanical coupling weight.

[0237] The attention score is determined based on the second dynamic feature data;

[0238] The second weight is determined based on the first weight and the attention score;

[0239] Based on the second dynamic feature data and the second weight, determine the state vector of each component;

[0240] Based on the second dynamic feature data and the state vectors of each component, a comprehensive representation of the global features is determined;

[0241] The second predicted lifetime is determined based on the comprehensive characterization of the global features and the second dynamic feature data.

[0242] Optionally, the second predicted lifetime is determined based on the comprehensive representation of the global features and the second dynamic feature data, including:

[0243] Based on the comprehensive representation of the global features and the second dynamic feature data, the hidden state vector is determined;

[0244] Based on the second dynamic feature data, determine the embedding vector for the degradation stage;

[0245] The hidden state vector is concatenated with the embedding vector of the degradation stage to obtain the concatenated vector;

[0246] The lifetime residual is determined based on the splicing vector.

[0247] The second predicted lifetime is determined based on the lifetime residual and the current lifetime.

[0248] Optionally, the first predicted lifetime and the second predicted lifetime are fused to obtain the predicted lifetime of the transformer, including:

[0249] The weights are determined based on the hidden state vector and the second static feature data.

[0250] The predicted life of the transformer is determined based on the first predicted life, the second predicted life, and the assigned weights.

[0251] It should be noted that this device is a device corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0252] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method of life prediction of a transformer, characterized in that, The method comprises the following steps: obtaining static characteristic data and dynamic characteristic data of a transformer; preprocessing the static characteristic data to obtain first static characteristic data; preprocessing the dynamic characteristic data to obtain first dynamic characteristic data; encoding the first static characteristic data to obtain second static characteristic data; encoding the first dynamic characteristic data to obtain second dynamic characteristic data; extracting features from the second dynamic characteristic data to obtain a first predicted life; extracting features from the second static characteristic data and the second dynamic characteristic data to obtain a second predicted life; fusing the first predicted life and the second predicted life to obtain a predicted life of the transformer; extracting features from the second dynamic characteristic data to obtain a first predicted life, comprising: According to , determining a first predicted life; wherein, RUL phy is a first predicted lifetime, L 0 is an initial lifetime, u is t 0- t is a transient time interval in the time instant, D is a temperature sensitivity coefficient, x u is real-time second dynamic characteristic data, x 0 is a reference second dynamic characteristic data;​ extracting features from the second static characteristic data and the second dynamic characteristic data to obtain a second predicted life, comprising: determining a first weight from the second dynamic characteristic data, the first weight being a thermal conduction weight, an electromagnetic coupling weight, or a mechanical coupling weight; determining an attention score from the second dynamic characteristic data; determining a second weight from the first weight and the attention score; determining a state vector of each component from the second dynamic characteristic data and the second weight; determining a comprehensive representation of a global feature from the second dynamic characteristic data and the state vector of each component; determining a second predicted life from the comprehensive representation of the global feature and the second dynamic characteristic data; determining a second predicted life from the comprehensive representation of the global feature and the second dynamic characteristic data, comprising: determining a hidden state vector from the comprehensive representation of the global feature and the second dynamic characteristic data; determining an embedding vector of a degradation stage from the second dynamic characteristic data; concatenating the hidden state vector and the embedding vector of the degradation stage to obtain a concatenated vector; determining a life residual from the concatenated vector; determining a second predicted life from the life residual and a current life; fusing the first predicted life and the second predicted life to obtain a predicted life of the transformer, comprising: determining an allocation weight from the hidden state vector and the second static characteristic data; determining a predicted life of the transformer from the first predicted life, the second predicted life, and the allocation weight.

2. The life prediction method of a transformer according to claim 1, characterized by, The static characteristic data comprises rated capacity, ambient temperature, ambient humidity, voltage grade, and / or insulation material; The dynamic characteristic data comprises winding leakage current, winding dielectric loss factor, insulation resistance absorption ratio, and / or polarization index.

3. The life prediction method of a transformer according to claim 1, characterized by, The preprocessing of the static characteristic data to obtain first static characteristic data comprises: performing outlier exclusion on the static characteristic data and the dynamic characteristic data to obtain the first static characteristic data.

4. The life prediction method of a transformer according to claim 1, characterized by, The preprocessing of the dynamic characteristic data to obtain first dynamic characteristic data comprises: performing outlier exclusion, missing value completion, and data enhancement on the dynamic characteristic data to obtain the first dynamic characteristic data.

5. The life prediction method of a transformer according to claim 1, characterized by, The encoding of the first dynamic characteristic data to obtain second dynamic characteristic data comprises: The first dynamic feature data is subjected to a sliding window to obtain a time-varying statistical quantity sequence; The time-varying statistical quantity sequence is transformed into a high-frequency energy proportion; All high-frequency energy proportions are summarized to obtain second dynamic feature data.

6. A life prediction device of a transformer characterized by comprising: Comprise: The acquisition module is used for acquiring the static feature data and dynamic feature data of the transformer; The processing module is used for pre-processing the static feature data to obtain first static feature data; pre-processing the dynamic feature data to obtain first dynamic feature data; feature encoding the first static feature data to obtain second static feature data; feature encoding the first dynamic feature data to obtain second dynamic feature data; feature extraction on the second dynamic feature data to obtain a first predicted life; feature extraction on the second static feature data and 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; The feature extraction on the second dynamic feature data to obtain a first predicted life comprises: According to , determining a first predicted life; wherein, RUL phy is a first predicted lifetime, L 0 is an initial lifetime, u is t 0- t is a transient time interval in the time instant, D is a temperature sensitivity coefficient, x u is real-time second dynamic characteristic data, x 0 is a reference second dynamic characteristic data;​ The feature extraction on the second static feature data and second dynamic feature data to obtain a second predicted life comprises: Determine a first weight according to the second dynamic feature data, wherein the first weight is a heat conduction weight, an electromagnetic coupling weight or a mechanical coupling weight; Determine an attention score according to the second dynamic feature data; Determine a second weight according to the first weight and the attention score; Determine a state vector of each component according to the second dynamic feature data and the second weight; Determine a comprehensive representation of a global feature according to the second dynamic feature data and the state vector of each component; Determine a second predicted life according to the comprehensive representation of the global feature and the second dynamic feature data; Determine a second predicted life according to the comprehensive representation of the global feature and the second dynamic feature data, comprising: Determine a hidden state vector according to the comprehensive representation of the global feature and the second dynamic feature data; Determine an embedding vector of a degradation stage according to the second dynamic feature data; Concatenate the hidden state vector and the embedding vector of the degradation stage to obtain a concatenated vector; Determine a life residual according to the concatenated vector; Determine a second predicted life according to the life residual and a current life; Fuse the first predicted life and the second predicted life to obtain the predicted life of the transformer, comprising: Determine an allocation weight according to the hidden state vector and the second static feature data; Determine the predicted life of the transformer according to the first predicted life, the second predicted life and the allocation weight.

Images