Intelligent prediction system for drying end point of transformer

Abstract

The invention belongs to the technical field of transformer insulation drying treatment and intelligent monitoring, and particularly provides an intelligent prediction system for a drying end point of a transformer, and the system comprises the steps: dividing a drying process into multiple stages through a K-means algorithm based on historical data and a physical mechanism, and outputting a feature baseline, the method comprises the following steps: constructing a multiple linear regression model of a basic threshold value and an adjustment variable, realizing personalized adaptation of the threshold value, accurately identifying a current stage through Euclidean distance matching of a real-time feature and a clustering feature, calculating a dynamic personalized threshold value in combination with a real-time time sequence feature and the adjustment variable, and embedding a feature-time two-dimensional attention layer in an LSTM (Long Short Term Memory) model. And focusing the key information based on a personalized threshold guide model, and outputting an end point prediction result. According to the method, staged intelligent control and terminal point accurate prediction of the drying process can be realized, the risk of insufficient or excessive drying is reduced, and different equipment and working conditions are adapted.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of transformer insulation drying treatment and intelligent monitoring, and particularly relates to a drying end point intelligent prediction system for a transformer. BACKGROUND

[0002] The drying treatment of a transformer is an extremely important link in the manufacturing process of power equipment, and has a direct impact on the insulation performance, service life and operation safety of the transformer. The quality of the internal insulation system of the transformer is crucial, and is a key factor to ensure the safe and reliable operation of the transformer under high voltage and high charge. The drying treatment is one of the basic links to ensure the quality of the insulation system during manufacturing and maintenance.

[0003] In the drying treatment of the transformer, accurate determination of the drying end point is the core of balancing drying quality and economy: determination too early will result in excessive moisture content of the insulation, causing partial discharge, insulation breakdown and other faults; determination too late will cause waste of energy consumption and accelerated aging of the insulation material.

[0004] The existing drying process control and end point determination technology has significant defects: first, the division of the drying stage relies on artificial experience to preset time nodes, such as forcibly switching to the pressure reduction stage after heating for 24 hours, ignoring individual differences such as equipment capacity, insulation type and initial moisture content, resulting in disconnection between stage boundaries and actual physical processes; second, the process control uses a one-size-fits-all fixed threshold, such as uniformly setting the water outflow rate <0.01 g / (kg h) in the high vacuum stage, which cannot adapt to the water migration law of different equipment and working conditions, and is prone to false alarms or omissions; third, the end point determination relies on the observation of index fluctuations by operation and maintenance personnel, lacks quantitative analysis of time series trends, is highly subjective and low in precision, and is difficult to avoid over-drying or insufficient drying. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a drying end point intelligent prediction system for a transformer.

[0006] The technical scheme adopted by the present application to solve the above technical problems is: A drying end point intelligent prediction system for a transformer, comprising the following modules: A feature clustering module: based on the historical drying data features of the transformer, the drying process is clustered, and the drying process is divided into multiple drying stages; A threshold mapping module: the key physical indicators of each drying stage are integrated into statistical features to form a stage feature baseline table, and the correlation analysis of the transformer state variables and the state determination threshold of each drying stage is performed to screen the adjustment variables, and a threshold mapping model of each drying stage is established according to the adjustment variables using multiple linear regression; Phase determination module: real-time monitoring of the characteristics of the current drying process, and matching analysis of the real-time monitoring characteristic data with the core characteristics of each drying phase obtained by clustering to determine the current belonging drying phase; Threshold setting module: for the real-time monitoring characteristics of the current belonging drying phase, the time sequence characteristics are extracted by using the sliding time window, and the adjusting variables are read in real time, which are substituted into the threshold mapping model to obtain the individualized threshold of the current transformer; Drying endpoint prediction module: deviation analysis of the individualized threshold of the drying phase and the real-time characteristics, calculation of the attention weight of the characteristic dimension and the time dimension, addition of the attention layer in the time sequence prediction model, and construction of the time sequence prediction model fused with the attention weight.

[0007] Further, the way of dividing the drying process into multiple drying phases is: Multiple core characteristics are obtained by screening and analyzing the historical drying data of the transformer; The elbow method is used to determine the number of clusters K, K starts from 1, the error sum of squares SSE corresponding to each K is calculated, the K-SSE curve is drawn, and the K corresponding to the point where the SSE drops sharply and then tends to be flat is the optimal value; Randomly select K core characteristics as initial cluster centers, calculate the Euclidean distance between each core characteristic and the K cluster centers, and assign the core characteristics to the cluster with the nearest Euclidean distance; After all core characteristics are assigned, the center of each cluster is recalculated, and the assignment and update are repeated until the cluster center variation is less than or equal to the preset threshold, and the clustering ends, finally obtaining multiple core characteristic groups; Each core characteristic group corresponds to a drying phase and a corresponding time interval, and the mean and standard deviation of the core characteristics of each drying phase are recorded.

[0008] Further, the way of obtaining the core characteristics is: Collect the historical drying data of multiple transformers of different models and different working conditions; The drying data of each transformer forms a time sequence sample, and each time sequence sample contains multiple time points; A fixed length sliding time window is set, statistical characteristics of the original data in each sliding time window are extracted, and each index in the time sequence sample is converted into a sample-feature matrix; Uniform the time stamps of all indexes to ensure that different dimension data in 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 characteristics of each index are screened, a candidate feature pool is constructed, and the statistical characteristics in the candidate feature pool are differentiated and ratio-converted, finally obtaining multiple core characteristics.

[0009] Further, the construction process of the stage feature baseline table is: For the multiple drying stages of clustering division, the mean and standard deviation of the core features in each drying stage are extracted, and based on the 3σ principle, the mean ± 3 times the standard deviation is taken as the feature interval of the core features of each drying stage to form a stage feature baseline table.

[0010] Further, the construction method of the threshold mapping model of each drying stage is: The lower limit of the feature interval of the core features of each drying stage in the stage feature baseline table is taken as the basic threshold of the drying stage. Based on any one drying stage: For the core features of the drying stage, the critical value of the determination index of each drying stage is obtained from the threshold library. Wherein, the threshold library is a threshold library for dividing the stages of the drying process of the transformer, and the critical value of the determination index is the critical value of the determination state of each drying stage. Select the variable that affects the threshold and can be quantified before drying, which is divided into two categories: equipment static variable and working condition initial variable, calculate the Pearson coefficient of each variable and threshold, and mark the variable that meets the requirement of the Pearson coefficient as the adjustment variable. The actual threshold corresponding to the critical value of the determination index is extracted from the historical drying data, and the corresponding adjustment variable is collected and quantitatively processed. The threshold mapping model is threshold = β0+ β1×X1+ β2×X2+...+ β n ×X n , wherein β0 is the reference threshold, X1~Xn is the adjustment variable, β1~β n is the coefficient, reflecting the influence weight of each adjustment variable on the threshold. The least square method with constraints is used to fit the model to minimize the error between the predicted value and the actual threshold. Finally, the threshold mapping model of each drying stage is obtained: threshold = basic threshold + ∑ (adjustment variable × coefficient).

[0011] Further, the judgment method of the current belonging drying stage is: The drying data of the current drying process of the transformer is monitored and collected in real time. The collected drying data is converted to obtain the core features corresponding to each core feature group. The Euclidean distance between the current core features and the core features of each drying stage obtained by clustering is calculated in real time, and the drying stage with the largest Euclidean distance is the current belonging drying stage.

[0012] Further, the calculation method of the individualized threshold of the current transformer is: The core features of the current drying stage are extracted from the stage feature baseline table, and the indexes contained in the core features are core monitoring indexes; The length of the sliding window is set according to the change rate of the key indexes, and the sliding step is synchronized with the real-time data acquisition frequency; The core monitoring indexes of the current drying stage are dynamically aggregated by using a sliding time window, and the real-time single-point data is converted into a time series feature vector; The inherent attributes and drying initial state parameters of the current transformer are called from the system database and used as the adjustment variable input of the threshold mapping model; The time series feature vector is input into the threshold mapping model of the corresponding drying stage, and the individualized threshold of the current transformer is calculated: individualized threshold = β0+Σ( time series featureᵢ × αᵢ ) + Σ( device static variable × βⱼ ) + Σ( initial working 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, γ k is the coefficient of the initial working condition variable, and all coefficients are obtained by training historical data.

[0013] Further, the construction process of the time series prediction model fused with attention weights is as follows: The feature dimension attention weight and the time dimension attention weight are calculated through real-time feature and individualized threshold deviation analysis; The feature dimension weight is weighted and summed with the input time series feature matrix to obtain the feature vector of the feature attention layer; The time dimension weight is weighted and summed with the hidden features output by the LSTM to obtain the feature vector of the time attention layer; The threshold deviation feature, the deviation change rate feature and the comprehensive matching degree are taken as the threshold correlation feature to form a feature matrix; The basic time series feature and the threshold correlation feature are spliced according to the time step to form the model total input matrix X (dimension: time step T × (F + 3)); The long short-term memory network LSTM is selected as the basic time series model, and the attention layer is embedded between the input layer and the hidden layer of the basic time series model, and the calculated feature dimension attention weight and time dimension attention weight are fused into the model; The overall architecture of the model is input layer -> feature attention layer -> LSTM layer -> time attention layer -> fully connected layer.

[0014] Further, the calculation method of the feature dimension attention weight is as follows: The real-time monitoring indexes of the last drying stage are extracted, and the statistical features are extracted by using a sliding window to form a basic time series feature matrix a normalized deviation of the core monitoring indicators from the personalized threshold value is calculated to obtain a threshold deviation feature; a deviation change rate of the threshold deviation feature, i.e., a sliding window change rate, is calculated: (current window deviation - previous window deviation) / window duration; a comprehensive matching degree is calculated according to the threshold deviation feature and the deviation change rate of the threshold deviation feature, comprehensive matching degree = ∑(ω i ×|threshold deviation feature i |), wherein ω i is a coefficient of each core monitoring indicator in the threshold mapping model; Pearson correlation coefficients r i of each core feature and the actual end time are calculated, and a weight W f0 (i) is calculated in combination with the comprehensive matching degree feature: weight W f0 (i) = |r i | x matching degree feature (i), wherein r i is the correlation coefficient of the ith core feature and the actual end.

[0015] Further, the calculation method of the time dimension attention weight is as follows: an exponential decay function is designed based on the deviation of the real-time feature and the personalized threshold value: W t0 (t) = exp(-k x D(t)), wherein D(t) is the absolute value of the threshold deviation feature at time t, and k is a decay function fitted from historical data.

[0016] The beneficial effects of the present application are as follows: through the feature engineering combining the K-means algorithm and the physical mechanism, the drying process is quantified into stages with clear physical meaning, replacing the artificial experience division, ensuring that the core objectives of each stage are clear, 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 stage-by-stage control; a threshold mapping model containing adjustment variables is constructed based on multiple linear regression, combining equipment static properties and initial working condition parameters, so that the threshold can dynamically adapt to transformers of different capacities, insulation types and moisture levels, reducing the false alarm rate of early warning; through the Euclidean distance matching of real-time features and clustered stage features, automatic recognition and dynamic updating of the drying stage are realized, ensuring that the subsequent modules such as threshold setting and end point prediction can accurately call the corresponding stage model and parameters, avoiding cross-stage misjudgment; in combination with the real-time time series features extracted by the sliding window and the updated adjustment variables, the personalized threshold is dynamically calculated, breaking through the limitations of static models, and can respond to working condition differences such as temperature fluctuations and vacuum degree changes in the drying process in a timely manner, improving the timeliness of the threshold; through the feature-time double dimension attention layer, the LSTM model is guided to focus on key information, the attention degree of features and time periods close to the end point is strengthened, the prediction error is reduced, and at the same time, the prediction result is interpretable through the attention weight matrix. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a step flow chart of the intelligent drying endpoint prediction system of the transformer of the present application; Figure 2 is a prediction flow chart of the time series prediction model fused with attention weights of the present application. DETAILED DESCRIPTION

[0018] The technical solutions of the present application will be described in further detail below in combination with the drawings and specific embodiments.

[0019] As shown in Figure 1 , the intelligent drying endpoint prediction system of the transformer provided by the embodiment of the present application comprises the following modules: a feature clustering module: based on the historical drying data features of the transformer, the drying process is clustered, and the drying process is divided into multiple drying stages; The process of dividing the drying process into multiple drying stages comprises: collecting historical drying data of multiple transformers of different models and different working conditions, including temperature field data, vacuum field data, moisture migration data, and process parameter data; Among them, the temperature field data includes winding average temperature, core temperature, oil tank wall temperature, and temperature standard deviation (axial / radial), reflecting the heating uniformity and heat conduction efficiency; The vacuum field data includes vacuum degree average, vacuum degree change rate, vacuum degree fluctuation amplitude, and pressure reduction rate (only for pressure reduction process), reflecting the vacuum system state; The moisture migration data includes unit time water discharge rate, cumulative water discharge amount, oil micro-water content (with oil drying), and water discharge rate change rate, reflecting the water discharge efficiency; The process parameter data includes heating power, medium spraying amount (gas phase drying), oil flow speed (hot oil circulation), and vacuum pump current, reflecting the process execution state; The drying data of 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 working conditions (initial moisture level, ambient temperature); Each continuous time series sample of each transformer is converted into a feature vector using a fixed-length sliding time window, specifically: Statistical features are extracted from the original data in each sliding time window, and each index in the time series sample is converted into a sample-feature matrix, for example: For the winding average temperature index, the mean value, maximum value, and change rate ((window end value-window initial value) / window length) in the window are extracted; For the vacuum degree index, the mean value, fluctuation amplitude (maximum value - minimum value), and stable duration (time continuously maintained in a certain interval) in the extraction window are extracted; The timestamps of all indicators are unified to ensure that different dimension data in the same sliding time window correspond to the same drying time, and are standardized to the [0, 1] interval; A candidate feature pool is constructed based on physical mechanisms, and only features that can reflect the differences in the core physical objectives of different drying stages are retained, for example: The core objective of the heating stage is uniform heating, and features such as temperature change rate and temperature standard deviation are retained; The core objective of the pressure reduction stage is stable vacuum pumping, and features such as vacuum degree change rate and pressure reduction rate are retained; The core objective of the high vacuum stage is deep dehydration, and features such as water removal rate and water removal rate change rate are retained; The original statistical features are differentiated and ratio-converted, for example: Derived temperature uniformity coefficient = temperature standard deviation / temperature mean (reflecting heating uniformity, this value is high in the heating stage and low in the high vacuum stage); Derived vacuum stability coefficient = 1 - vacuum fluctuation amplitude / vacuum mean (reflecting vacuum stability, this value is close to 1 in the high vacuum stage); Derived dehydration efficiency coefficient = unit time water removal rate / heating power (reflecting the relationship between energy consumption and dehydration effect, this value is high in the early stage of the high vacuum stage and low in the later stage); A number of core features are finally obtained; K-means algorithm is used to cluster the core features to divide the drying process into multiple drying stages, specifically: The elbow method is used to determine the number of clusters K, K starts from 1, the error sum of squares SSE corresponding to each K is calculated, the K-SSE curve is drawn, and the point corresponding to the optimal K where SSE drops sharply and then flattens is determined; 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 feature is assigned to the cluster with the closest Euclidean distance; After all core features are assigned, the center of each cluster is recalculated, and the assignment and update are repeated until the cluster center change 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 the corresponding time interval, and the core feature mean and standard deviation of each drying stage are recorded; The effect of dividing the drying process into multiple drying stages based on clustering is: Based on historical data and physical mechanisms, the continuous drying process is quantitatively divided into stages with clear physical meaning, laying a foundation for subsequent stage-by-stage control and stage-by-stage prediction.

[0020] Threshold mapping module: the statistical characteristics of the key physical indicators of each drying stage are integrated to form a stage characteristic baseline table, and the correlation analysis of the transformer state variables and the state determination threshold of each drying stage is performed to screen the adjustment variables, and a threshold mapping model of each drying stage is established by using multiple linear regression according to the adjustment variables; The integration process of the stage characteristic baseline table includes: For the multiple drying stages divided by clustering, the mean and standard deviation of the core features in each drying stage are extracted, and based on the 3σ principle, the mean ± 3 times the standard deviation is taken as the feature interval of each drying stage to form a stage characteristic baseline table, for example: | Drying stage | Core feature | Feature interval | Physical meaning |; | Heating stage | Temperature standard deviation | [0, 3] | Reflects the uniformity of the temperature field, and excessive large is easy to cause local overheating |; | Pressure reduction stage | Vacuum degree change rate | [500, 1500] | Reflects the stability of pressure reduction, and excessive large is easy to produce bubbles |; | High vacuum stage | Water output rate | [0.05, 0.2] | Reflects the dehydration efficiency, and excessive high may be accompanied by impurity precipitation |; The process of establishing a threshold mapping model of each drying stage by using multiple linear regression includes: Taking the lower limit of the normal interval of each drying stage core feature in the stage characteristic baseline table as the basic threshold of the drying stage; 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 properties and working condition differences, and the construction process needs to be designed in stages, specifically: Based on any drying stage: For the core features of the drying stage, the critical values of the determination indicators of each drying stage are obtained from the threshold library; Wherein, the threshold library is a threshold library for dividing the stages of the transformer drying process, and the critical value of the determination indicator is the critical value of triggering the early warning or determining the state of each drying stage, for example, the water output rate stable threshold of the high vacuum stage is the water output rate critical value when the insulation moisture content meets the standard; Select the variables that affect the threshold and can be quantified before / initial drying, which are divided into two categories: Equipment static variables: inherent properties of equipment, such as transformer capacity (MVA), insulation material type (such as corrugated paper / epoxy composite), winding structure (pie type / layer type), and operation time (reflecting aging degree); Initial working condition variables: state parameters at the start of drying, such as initial insulation moisture content (%), ambient temperature (℃), initial equipment temperature (℃), and moisture affected part (local / overall moisture affected); Screening basis: through correlation analysis (such as Pearson coefficient) and physical mechanism verification, variables with significant influence on threshold value are retained (such as the higher the initial moisture content, the more appropriate the water yield rate stable threshold needs to be improved); Extract the actual threshold value of each sample from historical drying data, which is determined by offline physical detection, for example, the water yield rate stable threshold label in the high vacuum stage is the water yield rate measured value when the insulation paper moisture content meets the standard (such as <0.5%); Collect the static variables of the equipment and the initial variables of the working conditions corresponding to the sample and perform quantitative processing; Divide the historical data into training set and test set in proportion; Use a multiple linear regression model to construct the threshold mapping model of each drying stage, specifically: Set the threshold mapping model as threshold = β0+ β1×X1+ β2×X2+...+ β n ×X n , where β0 is the baseline threshold, X1~Xn is the adjustment variable, β1~β n is the coefficient, reflecting the influence weight of each adjustment variable on the threshold; Use the least squares method with constraints to fit the model on the training set to minimize the error between the predicted value and the actual threshold value; Finally, the threshold mapping model of each drying stage is obtained: threshold = baseline threshold + ∑ (adjustment variable × coefficient); The role of constructing the threshold mapping model of each drying stage is: Based on the stage feature baseline obtained by clustering, a threshold model with general baseline + individual adjustment is constructed to solve the pain point that fixed threshold cannot adapt to different equipment and working conditions.

[0021] Stage determination module: real-time monitoring of the characteristics of the current drying process, and matching analysis of the real-time monitoring characteristic data and the core characteristics of each drying stage obtained by clustering to determine the current belonging drying stage; The process of matching analysis of the real-time monitoring characteristic data and the characteristics of each drying stage obtained by clustering to determine the current belonging drying stage includes: Real-time data acquisition through the online monitoring terminal of the transformer drying system (temperature sensor, vacuum degree transmitter, moisture detector, etc.), PLC control system and auxiliary monitoring equipment; The collection dimension must be completely consistent with the core feature set used in the clustering stage, and the collection frequency must be consistent with the sliding time window step length in the clustering; Use the standardized parameters trained in the clustering stage, that is, the mean and standard deviation of Z-score standardization, to convert the real-time characteristics: standardized characteristic value = (real-time characteristic - mean of clustering stage characteristics) / standard deviation of clustering stage characteristics; transforming the normalized real-time features into feature vectors in the format of the clustering phase 1; calculating the Euclidean distance between the real-time feature vector and the feature vector of each drying phase obtained by clustering, and the drying phase with the largest Euclidean distance is the current belonging drying phase; judging the role of the current belonging drying phase is: real-time monitoring the features of the current drying process, matching with the phase features obtained by clustering, accurately determining the current drying phase, and ensuring that the subsequent module calls the threshold value model and prediction logic corresponding to the phase; threshold setting module: for the real-time monitoring features of the current belonging drying phase, time series features are extracted by using a sliding time window, and the adjustment variables are read in real time, which are substituted into the threshold mapping model to obtain the individualized threshold value of the current transformer; the determination process of the individualized threshold value of the current transformer includes: extracting the core features of the current drying phase from the phase feature baseline table, and the indexes contained in the core features are the core monitoring indexes, which ensure that the input dimension of the threshold mapping model matches; using a sliding time window to dynamically aggregate the key indicators of the current phase, and transforming real-time single-point data into time series feature vectors to provide dynamic input for threshold calculation, specifically: setting the sliding window length according to the change rate of the key indicators, for example: indicators with fast changes have short windows, such as a 30-minute window for the change rate of vacuum degree; indicators with slow changes have long windows, such as a 2-hour window for water output rate; the sliding step is synchronized with the real-time data acquisition frequency; the inherent properties and drying initial state parameters of the current equipment are called from the system database as the adjustment variable input of the threshold mapping model, which ensures the individualized adaptation of threshold calculation, specifically: the adjustment variables include device static parameters and initial operating condition parameters, wherein the device static parameters are read from the transformer equipment archives (such as MES system or equipment management database), and the initial operating condition parameters are read from the detection records before the start of the drying process (such as offline detection report or process account book); inputting the time series feature vector and the adjustment variable into the threshold mapping model of the corresponding drying phase, and calculating the individualized threshold value of the current transformer through the pre-trained multiple linear regression formula: individualized threshold value = β0+Σ(time series featureᵢ×αᵢ)+Σ(device static variable×βⱼ)+Σ(initial operating condition variable k ×γ k ); wherein: β0 is the basic threshold value (from the phase feature baseline table), αᵢ is the coefficient of the time series feature; βⱼ is the coefficient of the device static variable, γ k is the coefficient of the initial operating condition variable (all coefficients are obtained by training historical data); The purpose of real-time calculation of personalized thresholds for each drying stage is: 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.

[0022] 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. like Figure 2 As shown, the construction process of the time-series prediction model that integrates attention weights includes: 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. 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. 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; 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; 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 (dimension: time step T×(F+3)). X is processed using the Z-score normalization parameters (mean μ, standard deviation σ) of the feature clustering module to ensure consistent input scale; Attention weights are divided into feature dimension weights and time dimension weights, and the calculation process includes: Feature dimension attention weight calculation: 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; Calculation of attention weights over time: Designing an exponential decay function based on real-time features and deviation of personalized threshold value, 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 fitted from historical data The long short-term memory network (LSTM) is selected as the basic time series model, and the attention layer is embedded between the input layer and the hidden layer of the basic time series model. The calculated feature dimension attention weight and time dimension attention weight are integrated into the model. The overall architecture of the model is input layer, feature attention layer, LSTM layer, time attention layer, and fully connected layer: Input: Time series matrix X (time step T x feature number F) Feature attention layer: Weighted feature matrix, highlighting high weight features LSTM layer: Extract time series dependent features of the feature matrix through 2-layer LSTM Time attention layer: Weighted time series features Fully connected layer: Map the weighted time series features to the end point prediction time through 2-layer fully connected layer Output: Predicted drying end point time, attention weight matrix The calculation logic of the feature attention layer is: The feature dimension weight W f is multiplied by the input time series feature matrix X to perform feature-by-feature weighted summation: The feature vector at time t in the time series feature matrix X is X(t)=[x1(t),x2(t),...,x F (t)], then the weighted feature vector X f (t)=[W f (1)×x1(t),W f (2)×x2(t),...,W f (F)×x F (t)]; The calculation logic of the time attention layer is: The time dimension weight W t is multiplied by the hidden feature H output by the LSTM to perform time-by-time weighted summation: 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); The role of constructing a time series prediction model that integrates attention weights is: Based on real-time timing characteristics and personalized threshold, through timing model + attention mechanism, the drying end point of high vacuum stage is accurately predicted, and the problem of over-drying or insufficient drying caused by relying on experience judgment is solved.

[0023] The drying end point intelligent prediction system of the transformer of the embodiment, based on the historical drying data characteristics of the transformer, clusters the drying process, divides the drying process into multiple drying stages, integrates the key physical indicators of each drying stage into statistical characteristics to form a stage feature baseline table, and analyzes the correlation between the transformer state variables and the state determination threshold of each drying stage to screen the adjustment variables;According to the adjustment variables, a threshold mapping model of each drying stage is established by using multiple linear regression, the characteristics of the current drying process are monitored in real time, and the real-time monitoring feature data is matched and analyzed with the core features of each drying stage obtained by clustering to determine the current drying stage;For the real-time monitoring features of the current drying stage, time sequence features are extracted by using a sliding time window, and the adjustment variables are read in real time, and the personalized threshold of the current transformer is obtained by substituting the threshold mapping model;The deviation analysis is performed on the personalized threshold of the drying stage and the real-time features, the attention weights of the feature dimension and the time dimension are calculated, the attention layer is added in the timing prediction model, and the timing prediction model integrated with the attention weight is constructed.

[0024] The present application divides the drying process into multiple stages by K-means algorithm and outputs feature baseline, constructs multiple linear regression model of basic threshold and adjustment variable, realizes threshold personalization adaptation, accurately identifies the current stage through the Euclidean distance matching of real-time features and clustering features, calculates dynamic personalized threshold combined with real-time timing characteristics and adjustment variables, embeds feature-time double-dimensional attention layer in LSTM model, and guides the model to focus on key information based on personalized threshold, and outputs the end point prediction result. The present application realizes the staged intelligent control and accurate prediction of the drying process, reduces the risk of insufficient drying or over-drying, and adapts to different equipment and working conditions.

[0025] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art can make various changes and substitutions to the technical solutions of the present application without departing from the spirit and scope of the present application, and these changes and substitutions all fall within the scope of the present application. The scope of protection required by the present application is defined by the appended claims and their equivalents.

Claims

1. A transformer dry-out endpoint intelligent prediction system, characterized by, Comprise: Feature clustering module: based on the historical drying data features of the transformer, the drying process is clustered, and the drying process is divided into multiple drying stages; Threshold mapping module: the key physical indicators of each drying stage are integrated into statistical features to form a stage feature baseline table, and the correlation analysis of the transformer state variables and the state judgment threshold of each drying stage is carried out to screen the adjustment variables, and the threshold mapping model of each drying stage is established by using multiple linear regression according to the adjustment variables; Stage determination module: real-time monitoring of the characteristics of the current drying process, and matching analysis of the real-time monitoring feature data and the core features of each drying stage obtained by clustering to determine the current drying stage; Threshold setting module: for the real-time monitoring features of the current drying stage, the time sequence features are extracted by using the sliding time window, and the adjustment variables are read in real time, and the personalized threshold of the current transformer is obtained by substituting the threshold mapping model; Drying endpoint prediction module: deviation analysis of the personalized threshold of the drying stage and the real-time features, calculation of the attention weight of the feature dimension and the time dimension, addition of the attention layer in the time series prediction model, and construction of the time series prediction model fused with the attention weight.

2. The intelligent drying endpoint prediction system for transformer according to claim 1, wherein: The drying process is divided into multiple drying stages in the following manner: A plurality of core features are obtained by screening and analyzing the historical drying data of the transformer; The elbow method is used to determine the number of clusters K, K is valued from 1, the error sum of squares SSE corresponding to each K is calculated, a K-SSE curve is drawn, and the K corresponding to the point where the SSE drops sharply and then tends to be flat is the optimal value; 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 cluster with the closest Euclidean distance; After all core features are assigned, the center of each cluster is recalculated, and the assignment and update are repeated until the cluster center variation is less than or equal to a preset threshold, the clustering ends, and finally a plurality of core feature groups are obtained; Each core feature group corresponds to a drying stage and a corresponding time interval, and the core feature mean and standard deviation of each drying stage are recorded.

3. The intelligent drying endpoint prediction system for transformer according to claim 2, wherein: The core features are obtained in the following manner: Collect historical drying data of multiple transformers of different models and different working conditions; The drying data of each transformer forms a time series sample, and each time series sample contains multiple time points; A fixed-length sliding time window is set, statistical features are extracted from the original data in each sliding time window, and each index in the time series sample is converted into a sample-feature matrix; The timestamps of all indices are unified to ensure that different dimension data in the same sliding time window correspond to the same drying time, and are standardized to the [0, 1] interval; Based on the physical mechanism, the statistical features of each index are screened to construct a candidate feature pool, and the statistical features in the candidate feature pool are differentiated and ratio-converted to finally obtain multiple core features.

4. The intelligent drying endpoint prediction system for transformer according to claim 1, wherein: The construction process of the stage feature baseline table is: for a plurality of drying stages of cluster division, the mean value and standard deviation of the core features in each drying stage are extracted, and based on the 3σ principle, the mean value ± 3 times the standard deviation is taken as the feature interval of the core features of each drying stage, The stage feature baseline table is formed.

5. The intelligent drying end point prediction system of transformer according to claim 4, characterized in that: The threshold mapping model of each drying stage is constructed in the following manner: The lower limit of the feature interval of the core features of each drying stage in the stage feature baseline table is taken as the basic threshold of the drying stage; Based on any one drying stage: For the core features of the drying stage, the critical value of the judgment index of each drying stage is obtained from the threshold library; Wherein, the threshold library is a threshold library for dividing each stage 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 the variable that affects the threshold and can be quantified before drying, which is divided into two categories: equipment static variable and working condition initial variable, calculate the Pearson coefficient of each variable and threshold, and mark the variable that meets the requirement of Pearson coefficient as the adjustment variable; The actual threshold corresponding to the critical value of the judgment index is extracted from the historical drying data, and the corresponding adjustment variable is collected and quantized; The threshold mapping model is threshold = β0+ β1× X1+ β2× X2+... + β n × X n n, wherein β0is a reference threshold, X1~ Xnare adjustment variables, β1~ β n nare coefficients, reflecting the influence weight of each adjustment variable on the threshold; The least square method with constraints is used to fit the model to minimize the error between the predicted value and the actual threshold; Finally, the threshold mapping model of each drying stage is obtained: threshold = basic threshold + ∑(adjustment variable × coefficient).

6. The intelligent drying end point prediction system of transformer according to claim 1, characterized in that: The judgment method of the current belonging drying stage is: Real-time monitoring and data collection are performed on the drying data of the current transformer drying process; The collected drying data is converted to obtain the core features corresponding to each core feature group; The Euclidean distance between the current core features and the core features of each drying stage obtained by clustering is calculated in real time, and the drying stage with the largest Euclidean distance is the current belonging drying stage.

7. The intelligent drying end point prediction system of transformer according to claim 1, characterized in that: The calculation method of the personalized threshold of the current transformer is: The core features of the current drying stage are extracted from the stage feature baseline table, and the indexes contained in the core features are the core monitoring indexes; The length of the sliding window is set according to the change rate of the key index, and the sliding step is synchronized with the real-time data collection frequency; The sliding time window is used to dynamically aggregate the core monitoring indexes of the current drying stage, and the real-time single point data is converted into a time series feature vector; The inherent attributes and drying initial state parameters of the current transformer are retrieved from the system database as the adjustment variable input of the threshold mapping model; mapping the time-series feature vector and the conditioning variables into a threshold mapping model corresponding to the drying phase, by calculating the personalized threshold for the current transformer: personalized threshold = β0+∑(time-series featureᵢ×αᵢ)+∑(device static variables×βⱼ)+∑(initial operating condition variables k ×γ k ) Wherein: β0 is the basic threshold, αᵢ is the coefficient of time sequence characteristics, βⱼ is the coefficient of equipment static variable, γ k is the coefficient of initial working condition variable, all coefficients are obtained by training historical data.

8. The intelligent drying end point prediction system of transformer according to claim 1, characterized in that: The construction process of the time series prediction model fused with attention weight is: Through real-time feature and personalized threshold deviation analysis, the feature dimension attention weight and time dimension attention weight are calculated; The feature dimension weight is weighted and summed with the input time series feature matrix to obtain the feature vector of the feature attention layer. The time dimension weight is weighted and summed with the hidden features output by the LSTM at each time point to obtain a feature vector of the time attention layer; The threshold deviation feature, the deviation change rate feature, and the comprehensive matching degree are taken as threshold correlation features to form a feature matrix; The basic time sequence feature and the threshold correlation feature are spliced according to the time step to form a total input matrix X of the model, with a dimension of T time steps x (F+3); The long short-term memory network LSTM is selected as the basic time sequence model, an attention layer is embedded between the input layer and the hidden layer of the basic time sequence model, and the calculated feature dimension attention weight and the time dimension attention weight are integrated into the model; The overall architecture of the model is an input layer, a feature attention layer, an LSTM layer, a time attention layer, and a fully connected layer.

9. The intelligent prediction system for the drying end point of the transformer according to claim 8, characterized in that: The calculation method of the feature dimension attention weight is: Real-time monitoring indicators of the last drying stage are extracted, statistical features are extracted using a sliding window, and a basic time sequence feature matrix is formed; A normalized deviation of the core monitoring indicator and the individualized threshold is calculated to obtain a threshold deviation feature; The deviation change rate of the threshold deviation feature, that is, the sliding window change rate, is calculated: (current window deviation - previous window deviation) / window length; According to the threshold deviation feature and the deviation change rate of the threshold deviation feature, a comprehensive matching degree is calculated, comprehensive matching degree =∑(ω i ×|threshold deviation feature i |), wherein ω i is a coefficient of each core monitoring index in the threshold mapping model; calculating a Pearson correlation coefficient r of each core feature with respect to the actual end time i , in combination with the comprehensive matching degree feature, calculating a weight W f0 (i): the weight W f0 (i) = |r i | x the matching degree feature (i), wherein r i is the correlation coefficient of the ith core feature with respect to the actual end time.

10. The intelligent prediction system for the drying end point of the transformer according to claim 9, characterized in that: The time dimension attention weight is calculated in the following manner: designing an exponential decay function based on the deviation of real-time features and individualization from a threshold value, W t0 (t) = exp(-k x D(t)), where D(t) is the absolute value of the threshold deviation feature at time t, and k is a decay function fitted from historical data.

Images