A smart prediction system for the drying endpoint of a transformer

Abstract

This invention belongs to the field of transformer insulation drying treatment and intelligent monitoring technology. Specifically, it provides an intelligent prediction system for the drying endpoint of transformers, comprising: dividing the drying process into multiple stages based on historical data and physical mechanisms, outputting feature baselines using the K-means algorithm, constructing a multiple linear regression model with basic thresholds and adjustment variables to achieve personalized threshold adaptation, accurately identifying the current stage by matching real-time features and clustering features with Euclidean distance, calculating dynamic personalized thresholds by combining real-time time-series features and adjustment variables, embedding a feature-time dual-dimensional attention layer in the LSTM model, guiding the model to focus on key information based on personalized thresholds, and outputting the endpoint prediction result. This invention can achieve phased intelligent control and accurate endpoint prediction of the drying process, reduce the risk of under-drying or over-drying, and is adaptable to different equipment and operating conditions.

Description

Technical Field

[0001] This invention belongs to the field of transformer insulation drying treatment and intelligent monitoring technology, specifically an intelligent prediction system for the drying endpoint of a transformer. Background Technology

[0002] Drying is a crucial step in the manufacturing process of power equipment, directly impacting the transformer's insulation performance, service life, and operational safety. The quality of the transformer's internal insulation system is paramount, a key factor in ensuring its safe and reliable operation under high voltage and high charge conditions. Drying is one of the fundamental steps in manufacturing and maintenance to guarantee the quality of the insulation system.

[0003] In the drying process of transformers, the accurate determination of the drying endpoint is the core of balancing drying quality and economy: determining it too early will lead to excessive moisture content in the insulation, causing faults such as partial discharge and insulation breakdown; determining it too late will result in energy waste and accelerated aging of insulation materials.

[0004] Existing drying process control and endpoint determination technologies have significant drawbacks: First, the division of drying stages relies on manual experience and preset time nodes, such as forcibly switching to the depressurization stage after 24 hours of heating, ignoring individual differences such as equipment capacity, insulation type, and initial moisture content, resulting in a disconnect between stage boundaries and the actual physical process; Second, process control uses a one-size-fits-all fixed threshold, such as uniformly setting the water output rate of the high vacuum stage to <0.01g / (kg・h), which cannot adapt to the moisture migration patterns of different equipment and operating conditions, and is prone to false alarms or missed alarms; Third, endpoint determination relies on maintenance personnel to observe indicator fluctuations, lacks quantitative analysis of time-series trends, is highly subjective and has low accuracy, and is difficult to avoid over-drying or under-drying. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent prediction system for the drying endpoint of transformers.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] A smart prediction system for the drying endpoint of a transformer includes the following modules:

[0008] Feature clustering module: Based on the historical drying data features of transformers, the drying process is clustered and divided into multiple drying stages;

[0009] Threshold mapping module: The key physical indicators of each drying stage are statistically integrated to form a stage characteristic baseline table. Correlation analysis is performed on the transformer state variables and the state judgment thresholds of each drying stage. Adjustment variables are selected, and multiple linear regression is used to establish a threshold mapping model for each drying stage based on the adjustment variables.

[0010] Stage determination module: It monitors the characteristics of the current drying process in real time, and matches and analyzes the real-time monitored characteristic data with the core characteristics of each drying stage obtained by clustering to determine the current drying stage;

[0011] Threshold setting module: For the real-time monitoring characteristics of the current drying stage, a sliding time window is used to extract time-series features, and the adjustment variables are read in real time and substituted into the threshold mapping model to obtain the personalized threshold of the current transformer.

[0012] Drying endpoint prediction module: Performs deviation analysis on personalized thresholds and real-time features of the drying stage, calculates attention weights in the feature dimension and time dimension, adds an attention layer to the time series prediction model, and constructs a time series prediction model that integrates attention weights.

[0013] Furthermore, the method of dividing the drying process into multiple drying stages is as follows:

[0014] Several core characteristics were obtained through screening and analysis of historical drying data of transformers;

[0015] The number of clusters K is determined using the elbow method, with K starting from 1. The sum of squared errors (SSE) corresponding to each K is calculated, and the K-SSE curve is plotted. The optimal value of K is the point where the SSE drops sharply and then flattens out.

[0016] K core features are randomly selected as initial cluster centers. The Euclidean distance between each core feature and the K cluster centers is calculated, and the core features are assigned to the clusters with the closest Euclidean distance.

[0017] After all core features are assigned, the center of each cluster is recalculated. The assignment and update are repeated until the change in the cluster center is less than or equal to the preset threshold. The clustering ends and multiple core feature groups are finally obtained.

[0018] Each core feature group corresponds to a drying stage and a corresponding time interval, and the mean and standard deviation of the core features for each drying stage are recorded.

[0019] Furthermore, the core features are obtained as follows:

[0020] Historical drying data were collected from multiple transformers of different models and operating conditions.

[0021] The dry data of each transformer forms a time series sample, and each time series sample contains multiple time points;

[0022] Set a fixed-length sliding time window, extract statistical features from the raw data within each sliding time window, and transform each indicator in the time series sample into a sample-feature matrix;

[0023] Unify the timestamps of all indicators to ensure that data from different dimensions within the same sliding time window correspond to the same drying time, and standardize them to the [0,1] interval;

[0024] Based on the physical mechanism, the statistical features of each indicator are screened, a candidate feature pool is constructed, and the statistical features in the candidate feature pool are differentially and proportionally transformed to finally obtain multiple core features.

[0025] Furthermore, the process of constructing the stage feature baseline table is as follows:

[0026] For the multiple drying stages divided by clustering, the mean and standard deviation of the core features within each drying stage are extracted. Based on the 3σ principle, the mean ± 3 times the standard deviation is used as the feature interval of the core features of each drying stage to form a stage feature baseline table.

[0027] Furthermore, the threshold mapping model for each drying stage is constructed as follows:

[0028] The lower limit of the feature interval of the core feature of each drying stage in the stage feature baseline table is used as the basic threshold of the drying stage.

[0029] Based on any drying stage:

[0030] For the core characteristics of the drying stage, the critical values ​​of the judgment index for each drying stage are obtained from the threshold library;

[0031] Among them, the threshold library is a threshold library used to divide the various stages of the transformer drying process, and the critical value of the judgment index is the critical value of the judgment state of each drying stage.

[0032] Select variables that affect the threshold and are quantifiable before drying, and divide them into two categories: equipment static variables and initial operating condition variables. Calculate the Pearson coefficient of each variable with respect to the threshold, and mark the variables whose Pearson coefficients meet the requirements as adjustment variables.

[0033] Extract the actual threshold corresponding to the critical value of the judgment index from historical drying data, and collect the corresponding adjustment variables for quantitative processing;

[0034] The threshold mapping model is: threshold = β0 + β1 × X1 + β2 × X2 + ... + β n ×X n Where β0 is the baseline threshold, X1~Xn are adjustment variables, and β1~β n The coefficients represent the weights of each adjustment variable on the threshold.

[0035] A constrained least squares method is used to fit the model, minimizing the error between the predicted value and the actual threshold.

[0036] The final threshold mapping model for each drying stage is obtained: Threshold = Basic threshold + ∑ (Adjustment variable × coefficient).

[0037] Furthermore, the method for determining the current drying stage is as follows:

[0038] Real-time monitoring and data acquisition of drying data during the current transformer drying process;

[0039] The collected drying data is transformed to obtain the core features corresponding to each core feature group;

[0040] The Euclidean distance between the current core feature and the core features of each drying stage obtained by clustering is calculated in real time. The drying stage with the largest Euclidean distance is the current drying stage.

[0041] Furthermore, the personalized threshold of the current transformer is calculated as follows:

[0042] Extract the core features of the current drying stage from the stage feature baseline table. The indicators included in the core features are the core monitoring indicators.

[0043] The sliding window length is set according to the rate of change of key indicators, and the sliding step size is synchronized with the real-time data acquisition frequency.

[0044] A sliding time window is used to dynamically aggregate the core monitoring indicators of the current drying stage, transforming real-time single-point data into time-series feature vectors;

[0045] The inherent properties and initial drying state parameters of the current transformer are retrieved from the system database and used as the input of the adjustment variables for the threshold mapping model;

[0046] The time-series feature vector and the adjustment variable are input into the threshold mapping model corresponding to the drying stage. The personalized threshold of the current transformer is calculated as follows: Personalized threshold = β0 + Σ(time-series feature ᵢ × αᵢ) + Σ(equipment static variable × βⱼ) + Σ(initial operating condition variable) k ×γ k );

[0047] Where: β0 is the basic threshold, αᵢ is the coefficient of the time-series feature, βⱼ is the coefficient of the device static variable, and γ k These are the coefficients of the initial operating condition variables; all coefficients were obtained through training with historical data.

[0048] Furthermore, the construction process of the time-series prediction model that incorporates attention weights is as follows:

[0049] By analyzing the deviation between real-time features and personalized thresholds, the attention weights in the feature dimension and the time dimension are calculated.

[0050] The feature vector of the feature attention layer is obtained by performing a feature-by-feature weighted summation of the feature dimension weights and the input temporal feature matrix;

[0051] The feature vector of the temporal attention layer is obtained by weighting and summing the temporal dimension weights with the latent features output by the LSTM on a time-by-time basis.

[0052] The threshold deviation feature, the deviation change rate feature, and the comprehensive matching degree are used as threshold association features to form a feature matrix;

[0053] The basic time-series features and threshold-related features are concatenated by time step to form the total input matrix X of the model (dimension: time step T×(F+3)).

[0054] We selected the Long Short-Term Memory (LSTM) network as the basic temporal model and embedded an attention layer between the input layer and the hidden layer of the basic temporal model, integrating the calculated feature dimension attention weights and time dimension attention weights into the model.

[0055] The overall model architecture is as follows: input layer -> feature attention layer -> LSTM layer -> temporal attention layer -> fully connected layer.

[0056] Furthermore, the attention weights of the feature dimensions are calculated as follows:

[0057] Real-time monitoring indicators for the final drying stage were extracted, and statistical features were extracted using a sliding window to form a basic time-series feature matrix.

[0058] The normalized deviation between the core monitoring indicators and the personalized thresholds is calculated to obtain the threshold deviation characteristics;

[0059] Calculate the rate of change of the threshold deviation feature, i.e., the sliding window rate of change: (current window deviation - previous window deviation) / window duration;

[0060] Based on the threshold deviation characteristics and the rate of change of the threshold deviation characteristics, the comprehensive matching degree is calculated as follows: Comprehensive matching degree = ∑(ω i ×|Threshold bias characteristics i |), where ω i These are the coefficients of each core monitoring indicator in the threshold mapping model;

[0061] Calculate the Pearson correlation coefficient r between each core feature and the actual endpoint time. i Based on the comprehensive matching degree features, the weight W is calculated. f0 (i): Weight W f0 (i)=|r i |× Matching feature(i), where r i Let be the correlation coefficient between the i-th core feature and the actual endpoint.

[0062] Furthermore, the calculation method for the time dimension attention weight is as follows:

[0063] Based on the deviation between real-time features and personalization and threshold, an exponential decay function is designed, W t0 (t)=exp(-k×D(t)), where D(t) is the absolute value of the threshold deviation feature at time t, and k is the decay function, which is fitted from historical data.

[0064] The beneficial effects of this invention are as follows: Through feature engineering combining the K-means algorithm with physical mechanisms, the drying process is quantified into stages with clear physical meaning, replacing manual experience-based divisions and ensuring clear core objectives for each stage, such as focusing on uniform temperature rise in the heating stage and focusing on deep dehydration in the high-vacuum stage, providing an objective basis for phased control; a threshold mapping model with adjustment variables is constructed based on multiple linear regression, integrating static equipment attributes and initial operating parameters, enabling the threshold to dynamically adapt to transformers of different capacities, insulation types, and moisture levels, reducing false alarm rates; and through Euclidean distance matching between real-time features and clustering stage features, it achieves… Automatic identification and dynamic updating of the drying stage ensure that subsequent modules such as threshold setting and endpoint prediction can accurately call the corresponding stage's model and parameters, avoiding cross-stage misjudgment. Combining real-time temporal features extracted by the sliding window with updated adjustment variables, personalized thresholds are dynamically calculated, overcoming the limitations of static models and responding promptly to differences in operating conditions such as temperature fluctuations and vacuum changes during the drying process, thus improving the timeliness of thresholds. The LSTM model is guided to focus on key information through a feature-time dual-dimensional attention layer, strengthening the attention to features and time periods near the endpoint, reducing prediction errors, and the prediction results are interpretable through the attention weight matrix. Attached Figure Description

[0065] Figure 1 This is a flowchart of the steps of the intelligent prediction system for the drying endpoint of the transformer of the present invention;

[0066] Figure 2 This is a flowchart of the prediction process of the time-series prediction model that integrates attention weights according to the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0068] like Figure 1 As shown in the figure, an intelligent prediction system for the drying endpoint of a transformer provided in this embodiment of the invention includes the following modules:

[0069] Feature clustering module: Based on the historical drying data features of transformers, the drying process is clustered and divided into multiple drying stages;

[0070] The process of dividing the drying process into multiple drying stages includes:

[0071] Historical drying data were collected from multiple transformers of different models and operating conditions, including temperature field data, vacuum field data, moisture migration data, and process parameter data.

[0072] The temperature field data includes the average winding temperature, core temperature, tank wall temperature, and temperature standard deviation (axial / radial), reflecting heating uniformity and heat transfer efficiency.

[0073] Vacuum field data includes the average vacuum level, vacuum level change rate, vacuum level fluctuation amplitude, and pressure reduction rate (pressure reduction process only), reflecting the state of the vacuum system;

[0074] Moisture migration data includes water discharge rate per unit time, cumulative water discharge, trace water content in oil (drying with oil), and water discharge rate change rate, reflecting the efficiency of water removal.

[0075] Process parameter data include heating power, medium spraying volume (vapor phase drying), oil flow rate (hot oil circulation), and vacuum pump current, reflecting the process execution status;

[0076] The drying data for each transformer forms a time series sample, and each time series sample contains multiple time points to ensure that the data covers different transformer characteristics (capacity, insulation type) and operating conditions (initial moisture content, ambient temperature).

[0077] A fixed-length sliding time window is used to transform continuous time-series samples of each transformer into feature vectors, specifically:

[0078] Statistical features are extracted from the raw data within each sliding time window, and the indicators in the time series samples are transformed into a sample-feature matrix, for example:

[0079] For the winding average temperature index, extract the mean, maximum value, and rate of change within the window ((end value of window - initial value of window) / window duration).

[0080] For the vacuum level index, extract the mean, fluctuation range (maximum value - minimum value), and stability duration (time to remain continuously within a certain range) within the window.

[0081] Unify the timestamps of all indicators to ensure that data from different dimensions within the same sliding time window correspond to the same drying time, and standardize them to the [0,1] interval;

[0082] A candidate feature pool is constructed based on physical mechanisms, retaining only features that reflect the differences in the core physical objectives at different drying stages, such as:

[0083] The core objective of the heating stage is to achieve uniform temperature rise while preserving characteristics such as temperature change rate and temperature standard deviation.

[0084] The core objective of the depressurization phase is to smoothly pump vacuum while preserving characteristics such as the rate of change of vacuum degree and the depressurization rate.

[0085] The core objective of the high vacuum stage is deep dehydration, which requires preserving characteristics such as water output rate and water output rate change rate.

[0086] Perform differencing and ratio transformations on the original statistical characteristics, for example:

[0087] Derived temperature uniformity coefficient = temperature standard deviation / temperature mean (reflects heating uniformity; this value is high during the heating stage and low during the high vacuum stage).

[0088] Derivative vacuum stability coefficient = 1 - vacuum fluctuation amplitude / average vacuum value (reflects vacuum stability; this value is close to 1 in the high vacuum stage).

[0089] Dehydration efficiency coefficient = water output rate per unit time / heating power (reflects the relationship between energy consumption and dehydration effect; this value is high in the early stage of high vacuum and low in the later stage).

[0090] Ultimately, several core features were obtained;

[0091] The K-means algorithm is used to cluster the core features, dividing the drying process into multiple drying stages, specifically:

[0092] The number of clusters K is determined using the elbow method, with K starting from 1. The sum of squared errors (SSE) corresponding to each K is calculated, and the K-SSE curve is plotted. The optimal value of K is the point where the SSE drops sharply and then flattens out.

[0093] K core features are randomly selected as initial cluster centers. The Euclidean distance between each core feature and the K cluster centers is calculated, and the core features are assigned to the clusters with the closest Euclidean distance.

[0094] After all core features are assigned, the center of each cluster is recalculated. The assignment and update are repeated until the change in the cluster center is less than or equal to the preset threshold. The clustering ends and multiple core feature groups are finally obtained.

[0095] Each core feature group corresponds to a drying stage and the corresponding time interval, and the mean and standard deviation of the core features for each drying stage are recorded;

[0096] The purpose of dividing the drying process into multiple drying stages based on clustering is:

[0097] Based on historical data and physical mechanisms, the continuous drying process is quantitatively divided into stages with clear physical meaning, laying the foundation for subsequent staged control and staged prediction.

[0098] Threshold mapping module: The key physical indicators of each drying stage are statistically integrated to form a stage characteristic baseline table. Correlation analysis is performed on the transformer state variables and the state judgment thresholds of each drying stage. Adjustment variables are selected, and multiple linear regression is used to establish a threshold mapping model for each drying stage based on the adjustment variables.

[0099] The integration process of the stage feature baseline table includes:

[0100] For the multiple drying stages divided by clustering, the mean and standard deviation of the core features within each drying stage are extracted. Based on the 3σ principle, the mean ± 3 times the standard deviation is used as the feature interval for each drying stage, forming a stage feature baseline table, for example:

[0101] |Drying stage|Core characteristics|Characteristic intervals|Physical meaning|;

[0102] |Heating Stage|Temperature Standard Deviation|[0,3]|Reflects the uniformity of the temperature field; an excessively large deviation can easily lead to local overheating|;

[0103] |Depressurization Stage|Vacuum Degree Change Rate|[500,1500]|Reflects the stability of pressure reduction; excessively high rates can easily generate bubbles|;

[0104] |High vacuum stage|Water output rate|[0.05,0.2]|Reflects dehydration efficiency; excessively high rates may be accompanied by impurity precipitation|;

[0105] The process of establishing a threshold mapping model for each drying stage using multiple linear regression includes:

[0106] The lower limit of the normal range of the core features of each drying stage in the stage feature baseline table is used as the basic threshold of the drying stage.

[0107] It should be noted that the core of the threshold mapping model is to establish a personalized threshold calculation relationship of basic threshold + adjustment variable × coefficient, so that the threshold can adapt to different equipment attributes and operating conditions. The construction process needs to be designed in stages, specifically:

[0108] Based on any drying stage:

[0109] For the core characteristics of the drying stage, the critical values ​​of the judgment index for each drying stage are obtained from the threshold library;

[0110] Among them, the threshold library is a threshold library used to divide the various stages of the transformer drying process. The critical value of the judgment index is the critical value that triggers the warning or judgment state in each drying stage. For example, the water output rate stability threshold of the high vacuum stage refers to the critical value of water output rate when the insulation moisture content meets the standard.

[0111] The variables that affect this threshold and are quantifiable before / in the early stages of drying are selected and divided into two categories:

[0112] Static variables of equipment: inherent attributes of equipment, such as transformer capacity (MVA), insulation material type (e.g., crepe paper / epoxy composite material), winding structure (pancake / layer type), and service life (reflecting the degree of aging).

[0113] Initial variables of the operating conditions: State parameters at the start of drying, such as initial insulation moisture content (%), ambient temperature (°C), initial equipment temperature (°C), and damp areas (partial / overall dampness).

[0114] Screening criteria: Through correlation analysis (such as Pearson coefficient) and physical mechanism verification, variables that have a significant impact on the threshold are retained (such as the higher the initial water content, the more the stable water output threshold needs to be increased).

[0115] The actual thresholds for each sample are extracted from historical drying data and determined by offline physical testing. For example, the stable threshold label for water output during the high vacuum stage is the measured value of water output when the moisture content of the insulating paper meets the standard (e.g., <0.5%).

[0116] Collect the equipment static variables and initial operating condition variables of the corresponding samples and perform quantitative processing;

[0117] Historical data is divided into training and testing sets proportionally.

[0118] A threshold mapping model for each drying stage is constructed using a multiple linear regression model, specifically as follows:

[0119] The threshold mapping model is defined as threshold = β0 + β1 × X1 + β2 × X2 + ... + β n ×X n Where β0 is the baseline threshold, X1~Xn are adjustment variables, and β1~β n The coefficients represent the weights of each adjustment variable on the threshold.

[0120] The model is fitted on the training set using the constrained least squares method to minimize the error between the predicted value and the actual threshold.

[0121] The final threshold mapping model for each drying stage is obtained as follows: Threshold = Basic threshold + ∑ (Adjustment variable × coefficient);

[0122] The purpose of constructing threshold mapping models for each drying stage is as follows:

[0123] Based on the stage feature baseline obtained by clustering, a threshold model with a general baseline and personalized adjustment is constructed to solve the pain point that fixed thresholds cannot be adapted to different equipment and working conditions.

[0124] Stage determination module: It monitors the characteristics of the current drying process in real time, and matches and analyzes the real-time monitored characteristic data with the core characteristics of each drying stage obtained by clustering to determine the current drying stage;

[0125] The process of matching and analyzing the real-time monitored feature data with the features of each drying stage obtained from clustering to determine the current drying stage includes:

[0126] Data is collected in real time through the online monitoring terminal (temperature sensor, vacuum transmitter, moisture detector, etc.), PLC control system and auxiliary monitoring equipment of the transformer drying system;

[0127] The data collection dimensions must be completely consistent with the core feature set used in the clustering stage, and the data collection frequency must be consistent with the sliding time window step size during clustering.

[0128] The real-time features are transformed using the standardized parameters trained during the clustering phase, namely the Z-score standardized mean and standard deviation: Standardized feature value = (real-time feature - mean of clustering phase features) / standard deviation of clustering phase features;

[0129] The standardized real-time features are transformed into feature vectors in the same format as those in the clustering stage.

[0130] The Euclidean distance between the feature vector and the feature vector of each drying stage obtained by clustering is calculated in real time. The drying stage with the largest Euclidean distance is the current drying stage.

[0131] The purpose of determining the current drying stage is:

[0132] Real-time monitoring of the characteristics of the current drying process, matching with the stage characteristics obtained by clustering, accurately determining the current drying stage, and ensuring that subsequent module calls = the threshold model and prediction logic of the corresponding stage;

[0133] Threshold setting module: For the real-time monitoring characteristics of the current drying stage, a sliding time window is used to extract time-series features, and the adjustment variables are read in real time and substituted into the threshold mapping model to obtain the personalized threshold of the current transformer.

[0134] The process of determining the personalized threshold of the current transformer includes:

[0135] Extract the core features of the current drying stage from the stage feature baseline table. The indicators contained in the core features are the core monitoring indicators, ensuring that they match the input dimensions of the threshold mapping model.

[0136] A sliding time window is used to dynamically aggregate key indicators for the current stage, transforming real-time single-point data into time-series feature vectors to provide dynamic input for threshold calculation. Specifically:

[0137] The sliding window length is set according to the rate of change of key indicators. For example, indicators that change rapidly have a short window, such as the vacuum rate of change, which uses a 30-minute window; indicators that change slowly have a long window, such as the water output rate, which uses a 2-hour window.

[0138] The sliding step size is synchronized with the real-time data acquisition frequency;

[0139] The inherent attributes and initial drying state parameters of the current equipment are retrieved from the system database and used as inputs to the threshold mapping model to ensure personalized adaptation of the threshold calculation. Specifically:

[0140] The adjustment variables include equipment static parameters and initial operating condition parameters. The equipment static parameters are read from the transformer equipment archive (such as the MES system or equipment management database), and the initial operating condition parameters are read from the inspection records before the drying process is started (such as offline inspection reports or process ledgers).

[0141] The time-series feature vector and the adjustment variable are input into the threshold mapping model corresponding to the drying stage. The personalized threshold of the current transformer is calculated using a pre-trained multiple linear regression formula: Personalized threshold = β0 + Σ(time-series feature ᵢ × αᵢ) + Σ(equipment static variable × βⱼ) + Σ(initial operating condition variable) k ×γ k );

[0142] Where: β0 is the base threshold (from the stage feature baseline table), αᵢ is the coefficient of the time-series feature; βⱼ is the coefficient of the device static variable, γ k The coefficients of the initial operating condition variables (all coefficients were obtained through training with historical data).

[0143] The purpose of real-time calculation of personalized thresholds for each drying stage is:

[0144] Based on the stage determination results, the threshold mapping model of the corresponding stage is invoked, and the dynamic personalized threshold of the current device is calculated by combining real-time time series features and device / operating condition parameters.

[0145] Drying endpoint prediction module: Performs deviation analysis on personalized thresholds and real-time features of the drying stage, calculates attention weights in the feature dimension and time dimension, adds an attention layer to the time series prediction model, and constructs a time series prediction model that integrates attention weights.

[0146] like Figure 2 As shown, the construction process of the time-series prediction model that integrates attention weights includes:

[0147] Real-time monitoring indicators for the last drying stage are extracted from the threshold setting module, and statistical features are extracted using a sliding window to form a basic time series feature matrix.

[0148] Calculate the threshold deviation feature, which is the normalized deviation between the real-time feature and the personalized threshold, reflecting the physical distance between the current state and the endpoint.

[0149] Calculate the rate of change of the threshold deviation feature, i.e. the sliding window rate of change: (current window deviation - previous window deviation) / window duration, which reflects the speed of approaching / moving away from the endpoint;

[0150] Based on the deviation between multiple core monitoring indicators and their corresponding personalized thresholds, the overall matching degree is calculated as follows: Overall matching degree = ∑(ω i ×|Threshold bias characteristics i |, where ω i These are the coefficients of each index in the threshold mapping model;

[0151] The threshold deviation feature, the deviation change rate feature, and the comprehensive matching degree are used as threshold association features to form a feature matrix;

[0152] The basic time-series features and threshold-related features are concatenated by time step to form the total input matrix X of the model (dimension: time step T×(F+3)).

[0153] X is processed using the Z-score normalization parameters (mean μ, standard deviation σ) of the feature clustering module to ensure consistent input scale;

[0154] Attention weights are divided into feature dimension weights and time dimension weights, and the calculation process includes:

[0155] Feature dimension attention weight calculation:

[0156] Calculate the Pearson correlation coefficient r between each core feature and the actual endpoint time. i Based on the comprehensive matching degree features, the weight is calculated as follows: Weight W f0 (i)=|r i |× Matching feature(i), where r i Let be the correlation coefficient between the i-th core feature and the actual endpoint;

[0157] Attention weight calculation in the time dimension:

[0158] An exponential decay function, W, is designed based on the deviation between real-time features and personalized thresholds. t0 (t) = exp(-k×D(t)), where D(t) is the absolute value of the threshold deviation feature at time t, and k is the decay function fitted from historical data.

[0159] Long Short-Term Memory (LSTM) network is selected as the basic time series model. An attention layer is embedded between the input layer and the hidden layer of the basic time series model, and the calculated attention weights of the feature dimension and the time dimension are integrated into the model.

[0160] The overall model architecture consists of an input layer, a feature attention layer, an LSTM layer, a temporal attention layer, and a fully connected layer.

[0161] Input: Time series matrix X (time step T × number of features F);

[0162] Feature attention layer: a weighted feature matrix that highlights high-weight features;

[0163] LSTM layer: Extracts temporal dependency features of the feature matrix through two LSTM layers;

[0164] Temporal attention layer: weighted temporal features;

[0165] Fully connected layer: Two fully connected layers are used to map the weighted temporal features to the predicted endpoint time;

[0166] Output: Estimated drying end time, attention weight matrix;

[0167] The computational logic of the feature attention layer is as follows:

[0168] Weight W of feature dimension f The eigenvectors of the input time-series feature matrix X are weighted and summed feature by feature: the eigenvectors at time t in the time-series feature matrix X are X(t) = [x1(t), x2(t), ..., x...]. F [(t)], then the weighted eigenvector X f (t)=[W f (1)×x1(t),W f (2)×x2(t),...,W f (F)×x F (t)];

[0169] The computational logic of the temporal attention layer is as follows:

[0170] Weight the time dimension W t The hidden features H output by the LSTM are weighted and summed at each time step: the hidden feature vector at time t in the hidden feature H is H(t) = [h1(t), h2(t), ..., h D (t)] (D is the number of hidden units), then the weighted hidden feature H t =∑(W t (t)×H(t))(t=1~T;

[0171] The purpose of constructing a time-series prediction model that incorporates attention weights is:

[0172] Based on real-time time-series features and personalized thresholds, the drying endpoint of the high-vacuum stage is accurately predicted through a time-series model and attention mechanism, solving the problem of over-drying or under-drying caused by relying on experience-based judgment.

[0173] The intelligent prediction system for the drying endpoint of transformers in this embodiment clusters the drying process based on the historical drying data characteristics of the transformer, dividing the drying process into multiple drying stages. It integrates the key physical indicators of each drying stage using statistical features to form a stage feature baseline table. Furthermore, it analyzes the correlation between the transformer's state variables and the state judgment thresholds of each drying stage, selecting moderating variables. Based on the moderating variables, it uses multiple linear regression to establish a threshold mapping model for each drying stage. It monitors the characteristics of the current drying process in real time and matches the real-time monitored feature data with the core features of each drying stage obtained from clustering to determine the current drying stage. For the real-time monitored features of the current drying stage, it uses a sliding time window to extract time-series features and reads the moderating variables in real time, substituting them into the threshold mapping model to obtain the personalized threshold for the current transformer. It performs deviation analysis between the personalized threshold of the drying stage and the real-time features, calculates the attention weights of the feature dimension and the time dimension, adds an attention layer to the time-series prediction model, and constructs a time-series prediction model that integrates attention weights.

[0174] This invention divides the drying process into multiple stages using the K-means algorithm and outputs feature baselines. It constructs a multiple linear regression model with a base threshold and moderating variables to achieve personalized threshold adaptation. By matching the Euclidean distance between real-time features and clustering features, it accurately identifies the current stage. Combining real-time temporal features and moderating variables, it calculates a dynamic personalized threshold. An LSTM model embeds a feature-time dual-dimensional attention layer, guiding the model to focus on key information based on the personalized threshold, and outputs the endpoint prediction result. This invention achieves phased intelligent control and accurate endpoint prediction of the drying process, reducing the risk of under-drying or over-drying, and is adaptable to different equipment and operating conditions.

[0175] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art can make various changes and substitutions to the technical solutions of the present invention without departing from the spirit and scope of the invention, and all such changes and substitutions fall within the scope of the present invention as claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart prediction system for the drying endpoint of a transformer, characterized in that, include: Feature clustering module: Based on the historical drying data features of transformers, the drying process is clustered and divided into multiple drying stages; Threshold mapping module: The key physical indicators of each drying stage are statistically integrated to form a stage characteristic baseline table. Correlation analysis is performed on the transformer state variables and the state judgment thresholds of each drying stage. Adjustment variables are selected, and multiple linear regression is used to establish a threshold mapping model for each drying stage based on the adjustment variables. Stage determination module: It monitors the characteristics of the current drying process in real time, and matches and analyzes the real-time monitored characteristic data with the core characteristics of each drying stage obtained by clustering to determine the current drying stage; Threshold setting module: For the real-time monitoring characteristics of the current drying stage, a sliding time window is used to extract time-series features, and the adjustment variables are read in real time and substituted into the threshold mapping model to obtain the personalized threshold of the current transformer. Drying endpoint prediction module: Performs deviation analysis on personalized thresholds and real-time features of the drying stage, calculates attention weights in the feature dimension and time dimension, adds an attention layer to the time series prediction model, and constructs a time series prediction model that integrates attention weights. The construction process of the stage feature baseline table is as follows: for multiple drying stages divided by clustering, the mean and standard deviation of the core features in each drying stage are extracted. Based on the 3σ principle, the mean ± 3 times the standard deviation is used as the feature interval of the core features of each drying stage to form the stage feature baseline table. The threshold mapping model for each drying stage is constructed as follows: The lower limit of the feature interval of the core feature of each drying stage in the stage feature baseline table is used as the basic threshold of the drying stage. Based on any drying stage: For the core characteristics of the drying stage, the critical values ​​of the judgment index for each drying stage are obtained from the threshold library; Among them, the threshold library is a threshold library used to divide the various stages of the transformer drying process, and the critical value of the judgment index is the critical value of the judgment state of each drying stage. Select variables that affect the threshold and are quantifiable before drying, and divide them into two categories: equipment static variables and initial operating condition variables. Calculate the Pearson coefficient of each variable with respect to the threshold, and mark the variables whose Pearson coefficients meet the requirements as adjustment variables. Extract the actual threshold corresponding to the critical value of the judgment index from historical drying data, and collect the corresponding adjustment variables for quantitative processing; The threshold mapping model is: threshold = β0 + β1 × X1 + β2 × X2 + ... + β n ×X n Where β0 is the baseline threshold, X1~Xn are adjustment variables, and β1~β n The coefficients represent the weights of each adjustment variable on the threshold. A constrained least squares method is used to fit the model, minimizing the error between the predicted value and the actual threshold. The final threshold mapping model for each drying stage is obtained as follows: Threshold = Basic threshold + ∑ (Adjustment variable × coefficient); The construction process of the time series prediction model that incorporates attention weights is as follows: By analyzing the deviation between real-time features and personalized thresholds, the attention weights in the feature dimension and the time dimension are calculated. The feature vector of the feature attention layer is obtained by performing a feature-by-feature weighted summation of the feature dimension weights and the input temporal feature matrix; The feature vector of the temporal attention layer is obtained by weighting and summing the temporal dimension weights with the latent features output by the LSTM on a time-by-time basis. The threshold deviation feature, the deviation change rate feature, and the comprehensive matching degree are used as threshold association features to form a feature matrix; The basic time-series features and threshold-related features are concatenated by time step to form the total input matrix X of the model, with the dimension being time step T×(F+3). We selected the Long Short-Term Memory (LSTM) network as the basic temporal model and embedded an attention layer between the input layer and the hidden layer of the basic temporal model, integrating the calculated feature dimension attention weights and time dimension attention weights into the model. The overall architecture of the model consists of an input layer, a feature attention layer, an LSTM layer, a temporal attention layer, and a fully connected layer. The attention weights for the aforementioned feature dimensions are calculated as follows: Real-time monitoring indicators for the final drying stage were extracted, and statistical features were extracted using a sliding window to form a basic time-series feature matrix. The normalized deviation between the core monitoring indicators and the personalized thresholds is calculated to obtain the threshold deviation characteristics; Calculate the rate of change of the threshold deviation feature, i.e., the sliding window rate of change: (current window deviation - previous window deviation) / window duration; Based on the threshold deviation characteristics and the rate of change of the threshold deviation characteristics, the comprehensive matching degree is calculated as follows: Comprehensive matching degree = ∑(ω i ×|Threshold bias characteristics i |), where ω i These are the coefficients of each core monitoring indicator in the threshold mapping model; Calculate the Pearson correlation coefficient r between each core feature and the actual endpoint time. i Based on the comprehensive matching degree features, the weight W is calculated. f0 (i): Weight W f0 (i)=|r i |× Matching feature(i), where r i Let be the correlation coefficient between the i-th core feature and the actual endpoint; The time-dimensional attention weight is calculated as follows: an exponential decay function, W, is designed based on the deviation between real-time features, personalization, and the threshold. t0 (t)=exp(-k×D(t)), where D(t) is the absolute value of the threshold deviation feature at time t, and k is the decay function, which is fitted from historical data.

2. The intelligent prediction system for the drying endpoint of a transformer according to claim 1, characterized in that: The method of dividing the drying process into multiple drying stages is as follows: Several core characteristics were obtained through screening and analysis of historical drying data of transformers; The number of clusters K is determined using the elbow method, with K starting from 1. The sum of squared errors (SSE) corresponding to each K is calculated, and the K-SSE curve is plotted. The optimal value of K is the point where the SSE drops sharply and then flattens out. K core features are randomly selected as initial cluster centers. The Euclidean distance between each core feature and the K cluster centers is calculated, and the core features are assigned to the clusters with the closest Euclidean distance. After all core features are assigned, the center of each cluster is recalculated. The assignment and update are repeated until the change in the cluster center is less than or equal to the preset threshold. The clustering ends and multiple core feature groups are finally obtained. Each core feature group corresponds to a drying stage and a corresponding time interval, and the mean and standard deviation of the core features for each drying stage are recorded.

3. The intelligent prediction system for the drying endpoint of a transformer according to claim 2, characterized in that: The core features are obtained as follows: Historical drying data were collected from multiple transformers of different models and operating conditions. The dry data of each transformer forms a time series sample, and each time series sample contains multiple time points; Set a fixed-length sliding time window, extract statistical features from the raw data within each sliding time window, and transform each indicator in the time series sample into a sample-feature matrix; Unify the timestamps of all indicators to ensure that data from different dimensions within the same sliding time window correspond to the same drying time, and standardize them to the [0,1] interval; Based on the physical mechanism, the statistical features of each indicator are screened, a candidate feature pool is constructed, and the statistical features in the candidate feature pool are differentially and proportionally transformed to finally obtain multiple core features.

4. The intelligent prediction system for the drying endpoint of a transformer according to claim 1, characterized in that: The method for determining the current drying stage is as follows: Real-time monitoring and data acquisition of drying data during the current transformer drying process; The collected drying data is transformed to obtain the core features corresponding to each core feature group; The Euclidean distance between the current core feature and the core features of each drying stage obtained by clustering is calculated in real time. The drying stage with the largest Euclidean distance is the current drying stage.

5. The intelligent prediction system for the drying endpoint of a transformer according to claim 1, characterized in that: The calculation method for the personalized threshold of the current transformer is as follows: Extract the core features of the current drying stage from the stage feature baseline table. The indicators included in the core features are the core monitoring indicators. The sliding window length is set according to the rate of change of key indicators, and the sliding step size is synchronized with the real-time data acquisition frequency. A sliding time window is used to dynamically aggregate the core monitoring indicators of the current drying stage, transforming real-time single-point data into time-series feature vectors; The inherent properties and initial drying state parameters of the current transformer are retrieved from the system database and used as the input of the adjustment variables for the threshold mapping model; The time-series feature vector and the adjustment variable are input into the threshold mapping model corresponding to the drying stage. The personalized threshold of the current transformer is calculated as follows: Personalized threshold = β0 + Σ(time-series feature ᵢ × αᵢ) + Σ(equipment static variable × βⱼ) + Σ(initial operating condition variable) k ×γ k ); Where: β0 is the basic threshold, αᵢ is the coefficient of the time-series feature, βⱼ is the coefficient of the device static variable, and γ k These are the coefficients of the initial operating condition variables; all coefficients were obtained through training with historical data.

Images