Alpine high-altitude box-type transformer condensation triggered discharge risk identification and dehumidification linkage control method and system

Abstract

The application provides a high-cold high-altitude box-type transformer condensation trigger discharge risk identification and dehumidification linkage control method and system, and relates to the technical field of power equipment state monitoring. The method comprises the following steps: acquiring multi-source heterogeneous monitoring data, constructing a basic data set through clock alignment and abnormal compensation; establishing a multi-physical field state equation of condensation, electric-thermal coupling and insulation degradation; training an initial digital twin intelligent agent using historical samples; calculating the prediction residual and the consistency of the multi-source evidence to evaluate the availability, and triggering self-evolution when it is not trusted; taking the untrusted sample and its neighborhood sample as the target domain, completing self-evolution update through domain adversarial transfer learning, freezing the bottom layer features, and fine-tuning the state fusion decision layer; the updated twin body outputs the insulation health degree and the condensation discharge risk value, generates a dehumidification linkage control instruction, and realizes closed-loop control. The problems of multi-source data fusion difficulty, lack of adaptability and credibility leading to dehumidification control failure are solved, and the high-cold high-altitude adaptability is improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and more specifically, to a method and system for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers and for dehumidification linkage. Background Technology

[0002] In high-altitude and frigid regions, the extreme temperature differences between day and night, thin air, and drastic humidity fluctuations make box-type transformers highly susceptible to condensation inside the cabinet due to poor sealing and the "breathing effect." This can lead to moisture absorption of insulation components, surface flashover, and potentially triggering discharge faults. Currently, dehumidification control for condensation in box-type transformers generally relies on single-point temperature and humidity sensors inside the cabinet, operating by starting and stopping heaters or dehumidifiers using fixed thresholds. Identification of the risk of condensation-triggered discharge is also largely based on empirical criteria of environmental parameters or periodic manual inspections, failing to incorporate multi-source information such as electrical quantities, equipment temperature, operational events, and system-level event chains into joint analysis. Such methods struggle to capture early signs of partial discharge induced by condensation, and dehumidification devices cannot be precisely controlled based on actual insulation risks. In harsh conditions at high altitudes and frigid regions, dehumidification often lags or operates ineffectively, resulting in frequent false alarms and missed alarms, threatening power supply reliability.

[0003] To enhance the ability to perceive equipment status, existing research has introduced digital twin technology into power equipment condition assessment, attempting to reflect the operational behavior of physical entities by constructing virtual models. However, current digital twin methods for box-type transformers are generally trained based on fixed historical operating conditions, failing to address prominent issues such as multi-source heterogeneity, temporal mismatch, and missing anomalies between continuous monitoring data and discrete event records in high-altitude and cold environments. Furthermore, after deployment, the models struggle to autonomously adjust to sudden environmental changes, new load patterns, or novel condensation temperature rise characteristics, lacking online quantitative assessment of model reliability and self-evolutionary updates based on multi-source evidence. This leads to the overflow of the reliability boundary in the output results of the twin model during operating condition drift, failing to provide continuous and reliable decision-making basis for condensation-triggered discharge risk identification and dehumidification linkage control. Therefore, there is an urgent need for a method for condensation-triggered discharge risk identification and dehumidification linkage control that can integrate multi-source heterogeneous data, possess environmental adaptability, and maintain reliability boundaries. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers, and to solve the problems of existing methods for identifying and controlling the risk of condensation discharge in high-altitude and cold-climate box-type transformers, which are difficult to integrate multi-source heterogeneous data, lack environmental adaptability and reliability assessment, resulting in delayed risk identification and dehumidification control failure.

[0005] This invention is achieved through the following technical solution: A method for identifying the risk of condensation-triggered discharge and linking it with dehumidification control in a high-altitude, cold-climate box-type transformer includes the following steps: Acquire multi-source heterogeneous monitoring data during the operation of box-type transformers, eliminate clock deviations between different data sources by aligning the multi-source heterogeneous monitoring data with a unified time reference, jointly compensate for data loss, sensor drift and communication delay anomalies, and construct a time-consistent basic observation dataset. Based on the basic observation dataset, a multi-physics field coupled state equation covering condensation formation, electro-thermal coupling temperature rise, and insulation moisture degradation is established. Using historical stable operation samples and historical condensation and discharge fault samples, discriminative features reflecting the risk of insulation degradation and partial discharge induced by condensation are trained and extracted to generate an initial twin intelligent agent and its state space expression that is dynamically mapped to physical entities. Based on the residual between the condensation-related state prediction value output by the twin agent and the actual monitoring quantity, the confidence interval and model availability index are calculated. When the prediction residual exceeds the preset boundary, the availability is lower than the set threshold, or the evidence consistency between electrical quantity, temperature quantity, event record and event chain is lower than the set threshold, the twin agent is determined to be in the untrusted area and self-evolution update is triggered. The untrusted samples and their temporal neighborhood samples are constructed as the target domain dataset, while historical stable operation samples, historical condensation and discharge fault samples, and calibrated trustworthy samples are constructed as the source domain dataset. The twin agent that triggers the update moment is used as the source domain model. The basic temporal feature extraction and event chain encoding capabilities of the source domain model are inherited through transfer learning. The domain adversarial feature adaptation method is used to reduce the distribution difference between the source domain model and the target domain in the fusion feature space. The bottom feature extraction layer is frozen, the state fusion layer and the output decision layer are fine-tuned, and step size constraints, convergence criteria and multi-source evidence consistency constraints are applied to make the model converge to the new trustworthy state boundary and complete the self-evolution update. The updated twin agent outputs insulation health and condensation-triggered discharge risk values ​​in real time. Based on the condensation-triggered discharge risk value and current environmental parameters, it generates dehumidification linkage control commands to control the start-stop and power adjustment of the heater or dehumidifier, realizing closed-loop linkage between condensation protection and discharge risk suppression.

[0006] Optionally, the unified time reference alignment is achieved by constructing a weighted fused time reference sequence, as shown in the following equation:

[0007] in, Represents a unified time reference series; Represents a Network Time Protocol (NTP) timestamp; Indicates a Precision Time Protocol (PTP) timestamp; Indicates satellite time synchronization timestamp; , and Let represent the weight coefficients of the corresponding time synchronization sources, and satisfy . , , , ; The weighting coefficients are determined by the standard deviation of the time deviation of each time source, as shown in the following formula:

[0008] in, Indicates the first The fusion weights of time-series sources; Indicates the first The standard deviation of the time deviation of the time source within the statistical window; time window resampling of data with different sampling frequencies based on a unified time reference sequence, as shown in the following formula:

[0009] in, This represents the data after resampling; Indicates the first time within the time window The original data of each sampling point; This indicates the number of sampling points within the time window.

[0010] Optionally, the joint compensation for data loss, sensor drift, and communication delay anomalies includes: For short-term missing data, an interpolation reconstruction method is used for recovery, as shown in the following formula:

[0011] in, Indicates missing moments The recovery value; and These represent the observations at both ends of the missing interval; For continuous states, Kalman smoothing is used for correction, and the state update expression is:

[0012]

[0013]

[0014]

[0015]

[0016] in, This represents the state estimate at the previous moment; This represents the estimated prior state value at the current moment; This represents the estimated posterior state at the current time. Represents the state transition matrix; Represents the control input matrix; Indicates external control input; Represents the prior error covariance matrix; Represents the posterior error covariance matrix; Represents the process noise covariance matrix; Indicates Kalman gain; Represents the observed value; Represents the observation matrix; Represents the observation noise covariance matrix; Represents the identity matrix; The state vector is set as shown in the following formula:

[0017] in, Indicates voltage status; Indicates the current state; Indicates power status; This indicates the oil temperature or the equivalent temperature of the winding. Indicates the ambient temperature status; Indicates the insulation health status; This indicates the event chain state code.

[0018] Optionally, the process noise covariance matrix and the observation noise covariance matrix are dynamically corrected in high-altitude and cold environments, as shown in the following equation:

[0019]

[0020] in, This represents the dynamically corrected process noise covariance matrix; Represents the reference process noise covariance matrix; This represents the dynamically corrected observation noise covariance matrix; Represents the reference observation noise covariance matrix; Indicates the change in ambient temperature; Indicates the change in altitude; Indicates the intensity of sensor fluctuations; This represents the communication delay statistics; , , and This is the corresponding empirical correction factor; A high-altitude environment correction factor is introduced to correct the heat transfer coefficient, as shown in the following formula:

[0021]

[0022] in, This represents the corrected heat transfer coefficient; This represents the baseline heat transfer coefficient under standard altitude conditions. This indicates the correction factor for high-altitude environments; Indicates altitude and ambient temperature air density below; This indicates the air density under standard atmospheric conditions. Indicates the heat exchange experience index; Indicates the ventilation or natural wind speed of the transformer substation; Indicates the reference wind speed; This represents the wind speed correction factor.

[0023] Optionally, the multiphysics coupled state equations include an electro-thermal coupled temperature rise model and an insulation state degradation model, the electro-thermal coupled temperature rise model being shown in the following equation:

[0024] in, This represents the heat capacity parameter; Indicates the temperature of key components; Indicates time; Indicates power loss; Indicates the heat transfer coefficient; Indicates the heat dissipation area; Indicates ambient temperature; Represents the flow field disturbance term; The insulation degradation model is shown in the following equation:

[0025]

[0026] in, express The level of insulation health at all times; Indicates the initial insulation health; Indicates the temperature rise degradation coefficient; Indicates the cumulative stress due to temperature rise; express Temperature of key components at any given moment; Indicates the reference temperature threshold for insulating materials; The state space representation of the twin agent is constructed as shown in the following equation:

[0027] in, express The state-space representation of a twin agent at any given moment; Indicates component-level features; Indicates loop-level characteristics; Indicates interval-level features; Indicates the characteristics of an event chain; This represents the state fusion mapping function.

[0028] Optionally, the calculation of the confidence interval includes: The model predicts residuals and constructs confidence boundaries, as shown in the following equation:

[0029] in, express The model prediction residuals at time step; Indicates the actual monitored value; Indicates the model's predicted value; The credibility boundary is shown in the following formula:

[0030] in, Indicates the credibility boundary; This represents the mean of the residuals; Indicates the standard deviation of the residuals; Represents the confidence coefficient; The confidence coefficient The dynamic adjustment is based on the residual distribution, as shown in the following formula:

[0031] in, express The dynamic confidence coefficient at time t; Indicates the baseline confidence coefficient; Represents the residual coefficient of variation; This represents the residual skewness or abnormal fluctuation statistic. Indicates the risk level of the event; , and This represents the corresponding weighting coefficient.

[0032] Optionally, the usability metrics of the computational model include: The model availability metric is established as shown in the following formula:

[0033] in, Indicates model availability; This indicates the number of samples that exceed the confidence boundary within the statistical window; This indicates the total number of samples within the statistics window; The consistency of the model output is analyzed under multiple internal perturbations, as shown in the following equation:

[0034] in, Indicates consistency index; Indicates the number of disturbance tests; Indicates the first Model output under the second perturbation; This represents the average of multiple perturbation outputs; Indicates the perturbation threshold; Indicates characteristic functions; When satisfied Exceeding the preset threshold Below the preset availability threshold or If any condition falls below the preset consistency threshold, the current twin agent is determined to be in an untrusted region and an update determination process is triggered.

[0035] Optionally, in the self-evolutionary update, the model updates the total loss function. Build as follows:

[0036] in, For prediction error loss; This represents a loss due to multi-source consistency. For regularization terms; Losses due to domain confrontation; The prediction error loss Calculate using the following formula:

[0037] in, This indicates the number of samples that participated in this update; Indicates the first The actual label, credible pseudo label, or historical confirmation status of each sample; The model represents the first The predicted output for each sample; The multi-source consistency loss Calculate using the following formula:

[0038] in, and These represent inputs from different data sources; and These represent the feature mapping functions of the corresponding data sources; Indicates data source With data source Consistency weights are used to constrain the consistency of state evidence among electrical quantities, temperature quantities, SOE events, and station-level event chains; The regularization term Calculate using the following formula:

[0039] in, This represents the currently updated model parameters; This represents the parameters of the source domain model before the update was triggered; The domain adversarial loss Calculate using the following formula:

[0040] in, Indicates a feature extractor; Representation domain discriminator; Indicates the first One source domain sample; Indicates the first One target domain sample; and These represent the number of samples in the source domain and the target domain, respectively.

[0041] Optionally, during the self-evolutionary update process, step size constraints are applied to the model parameters and a convergence criterion is established; Model parameter update increments are truncated using the following formula:

[0042] Model parameters are updated using the following formula:

[0043] in, Indicates the first Model parameters at the next update; Indicates the first Parameter increment during the next update; Indicates the learning rate; This represents the gradient of the total loss function with respect to the model parameters; Indicates the upper limit of the step size constraint; This indicates that the gradient update amount is limited to A cutoff function within the specified range; The upper limit of the step size constraint The target domain sample consistency is adaptively determined, as shown in the following formula:

[0044] in, Indicates the upper limit of the baseline step size; This represents the output consistency index of the target domain samples under multiple internal perturbations. Indicates the adjustment coefficient; The convergence criterion is shown in the following formula:

[0045] in, and They represent the first Next and first The loss value of the next iteration; This represents the convergence threshold. When the convergence criterion is met for multiple consecutive update cycles, and the updated model meets the requirements of residual boundary, consistency index, and model availability within the verification window, the twin agent is deemed to have completed a reliable self-evolutionary update.

[0046] Based on the same inventive concept, this invention also provides a risk identification and dehumidification linkage control system for condensation-triggered discharge of high-altitude and cold-climate box-type transformers, used to implement the aforementioned method for risk identification and dehumidification linkage control of condensation-triggered discharge of high-altitude and cold-climate box-type transformers, including: The multi-source data synchronization and compensation module is used to acquire multi-source heterogeneous monitoring data during the operation of the box-type transformer. It eliminates clock deviation by aligning with a unified time base, and performs joint compensation for data loss, sensor drift and communication delay anomalies, and outputs a basic observation dataset with consistent timing. The initialization module, connected to the multi-source data synchronization and compensation module, is used to receive the basic observation dataset, establish a multi-physics field coupled state equation covering condensation formation, electro-thermal coupling temperature rise and insulation moisture degradation, train and extract discriminative features using historical stable operation samples and historical condensation and discharge fault samples, and generate an initial twin intelligent agent and its state space expression that is dynamically mapped to the physical entity. The self-evolution trigger module, connected to the initialization module, is used to calculate the confidence interval and model availability index based on the residual between the condensation-related state prediction value output by the twin agent and the actual monitoring quantity, and to evaluate the consistency of multi-source evidence among electrical quantities, temperature quantities, event records and event chains. When the prediction residual exceeds the preset boundary, the availability is lower than the set threshold, or the evidence consistency is lower than the set threshold, the twin agent is determined to be in an untrustworthy area and a self-evolution trigger signal is issued. The domain adaptive update module, connected to the self-evolution trigger module, receives the self-evolution trigger signal, constructs the target domain dataset from samples deemed untrustworthy and their temporal neighborhood samples, and constructs the source domain dataset from historical stable operation samples, historical condensation and discharge fault samples, and calibrated trustworthy samples. The twin agent at the trigger update moment is used as the source domain model. Through transfer learning, it inherits the basic temporal feature extraction and event chain encoding capabilities of the source domain model. It adopts a domain adversarial feature adaptation method to reduce the distribution difference between the source domain model and the target domain in the fusion feature space, freezes the bottom feature extraction layer, fine-tunes the state fusion layer and the output decision layer, and applies step size constraints, convergence criteria and multi-source evidence consistency constraints to make the model converge to the new trustworthy state boundary, complete the self-evolution update, and output the updated twin agent parameters. The risk identification module, connected to the domain adaptive update module, is used to load the updated twin agent parameters, calculate and output the insulation health and condensation-triggered discharge risk values ​​in real time. The dehumidification linkage control module, connected to the risk identification module, is used to receive the condensation-triggered discharge risk value and current environmental parameters, generate dehumidification linkage control commands, and control the start / stop and power adjustment of the heater or dehumidifier to achieve closed-loop linkage between condensation protection and discharge risk suppression.

[0047] The technical solution of the present invention has at least the following advantages and beneficial effects: By aligning multi-source heterogeneous monitoring data with a unified time reference and jointly compensating for data loss, sensor drift, and communication delay anomalies, the temporal mismatch and inconsistency between continuous monitoring data and discrete event records in high-altitude and cold environments are effectively eliminated. A time-consistent basic observation dataset is constructed, which improves the data quality and reliability of subsequent analysis and modeling.

[0048] A multi-physics coupled state equation covering condensation formation, electro-thermal coupling temperature rise, and insulation moisture degradation is established. Discriminative features are extracted from historical stable operation and fault samples. The generated initial twin agent can deeply integrate multi-source information such as environmental quantities, electrical quantities, equipment temperature, and event chains. This overcomes the limitations of relying solely on single-point temperature and humidity monitoring within the cabinet and can capture early signs of insulation degradation and partial discharge induced by condensation earlier and more accurately.

[0049] By predicting residuals to calculate confidence intervals and model availability indicators, and combining the consistency of multi-source evidence between electrical quantities, temperature quantities, event records, and event chains, online quantitative assessment of credibility is conducted. When the twin agent enters an untrusted region, it automatically triggers self-evolutionary updates, solving the problem of traditional digital twin models overflowing the output credibility boundary when operating conditions drift. This ensures that the model maintains reliable risk identification capabilities under complex conditions such as environmental changes and new load patterns.

[0050] By employing transfer learning and domain adversarial feature adaptation to achieve self-evolutionary updates, the twin agent that triggers the update moment is used as the source domain model, inheriting its basic temporal feature extraction and event chain encoding capabilities. The state fusion layer and output decision layer are fine-tuned, and step size constraints, convergence criteria, and multi-source evidence consistency constraints are applied to enable the model to quickly converge to the new credible state boundary. This approach can fully utilize existing knowledge and effectively adapt to the differences in target domain data distribution, thus realizing the continuous autonomous evolution of the twin agent of the box transformer under varying operating conditions.

[0051] Based on the real-time updated insulation health and condensation-triggered discharge risk values ​​output by the twin agent, and combined with current environmental parameters, dehumidification linkage control commands are generated to dynamically adjust the start / stop and power of the heater or dehumidifier, achieving precise linkage between dehumidification control and actual insulation risk. Compared with fixed threshold start / stop methods, this avoids dehumidification lag or ineffective operation, reduces the probability of false alarms and missed alarms, and effectively suppresses insulation flashover and discharge faults caused by condensation in harsh operating conditions at high altitudes and cold regions, ensuring power supply reliability. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the high-altitude and cold-climate box-type transformer condensation-triggered discharge risk identification and dehumidification linkage control system according to an embodiment of the present invention. Detailed Implementation

[0053] The following is a detailed description of the embodiments, in conjunction with the accompanying drawings.

[0054] Reference Figure 1 A method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers, comprising the following steps: Step 1: Acquire multi-source heterogeneous monitoring data during the operation of the box-type transformer. Align the multi-source heterogeneous monitoring data with a unified time base to eliminate clock deviations between different data sources. Perform joint compensation for data loss, sensor drift, and communication delay anomalies to construct a time-consistent basic observation dataset. The multi-source heterogeneous monitoring data includes: continuously sampled electrical quantity and temperature data, discrete sequence event record (SOE) data, and station-level operation event chain data for wind turbines or inverters.

[0055] In some embodiments, the unified time reference alignment is achieved by constructing a weighted fused time reference sequence, as shown in the following equation:

[0056] in, Represents a unified time reference series; Represents a Network Time Protocol (NTP) timestamp; Indicates a Precision Time Protocol (PTP) timestamp; Indicates satellite time synchronization timestamp; , and Let represent the weight coefficients of the corresponding time synchronization sources, and satisfy . , , , This is used to ensure that the unified time reference sequence is a weighted fusion result from different time sources, avoiding scale shifts in the fused timestamps. In actual value selection, the weight coefficient ( , and The timing accuracy, packet loss rate, latency fluctuation, and availability of different time sources are determined; when the satellite timing signal is stable, the accuracy is improved. The value of ; when the precision time protocol link within the station is stable, improve . The value of ; when only network time synchronization is available, improve The value of .

[0057] The weighting coefficients are determined by the standard deviation of the time deviation of each time source, as shown in the following formula:

[0058] in, Indicates the first The fusion weights of time-series sources; Indicates the first The standard deviation of the time deviation of the time source within the statistical window; time window resampling of data with different sampling frequencies based on a unified time reference sequence, as shown in the following formula:

[0059] in, This represents the data after resampling; Indicates the first time within the time window The original data of each sampling point; This indicates the number of sampling points within the time window.

[0060] In some embodiments, the joint compensation for data loss, sensor drift, and communication delay anomalies includes: For short-term missing data, an interpolation reconstruction method is used for recovery, as shown in the following formula:

[0061] in, Indicates missing moments The recovery value; and These represent the observations at both ends of the missing interval; For continuous states, Kalman smoothing is used for correction, and the state update expression is:

[0062]

[0063]

[0064]

[0065]

[0066] in, This represents the state estimate at the previous moment; This represents the estimated prior state value at the current moment; This represents the estimated posterior state at the current time. Represents the state transition matrix; Represents the control input matrix; Indicates external control input; Represents the prior error covariance matrix; Represents the posterior error covariance matrix; Represents the process noise covariance matrix; Indicates Kalman gain; Represents the observed value; Represents the observation matrix; Represents the observation noise covariance matrix; The identity matrix is ​​represented. The state space dimension is determined based on the monitoring object of the box-type transformer. The state transition matrix... The observation matrix is ​​determined based on the state change relationship between adjacent sampling times; when precise mechanistic parameters are lacking, they are obtained through least squares identification using historical stable operating condition data; The process noise covariance matrix is ​​determined based on the actual observable variables of the sensor. If the sensor directly observes the corresponding component in the state vector, the corresponding position is set to 1; otherwise, it is set to 0. The observation noise covariance matrix is ​​determined by the sample covariance of the state prediction error under historical stable operating conditions. It is determined by the sensor accuracy, calibration error, and historical repeated measurement error.

[0067] Furthermore, the state vector is set as shown in the following equation:

[0068] in, Indicates voltage status; Indicates the current state; Indicates power status; This indicates the oil temperature or the equivalent temperature of the winding. Indicates the ambient temperature status; Indicates the insulation health status; This indicates the event chain state code.

[0069] In some embodiments, in high-altitude and cold environments, if low-temperature drift, increased communication delay, or changes in heat dissipation conditions due to low air pressure occur, the process noise covariance matrix and the observation noise covariance matrix are dynamically corrected, as shown in the following equation:

[0070]

[0071] in, This represents the dynamically corrected process noise covariance matrix; Represents the reference process noise covariance matrix; This represents the dynamically corrected observation noise covariance matrix; Represents the reference observation noise covariance matrix; Indicates the change in ambient temperature; Indicates the change in altitude; Indicates the intensity of sensor fluctuations; This represents the communication delay statistics; , , and The corresponding empirical correction coefficient can be obtained by calibration from historical operating data; A high-altitude environment correction factor is introduced to correct the heat transfer coefficient, as shown in the following formula:

[0072]

[0073] in, This represents the corrected heat transfer coefficient; This represents the baseline heat transfer coefficient under standard altitude conditions. This represents the high-altitude environment correction factor, used to characterize the combined effects of increased altitude, decreased air density, low temperature environment, and changes in wind-cooling efficiency on heat dissipation capacity. Indicates altitude and ambient temperature air density below; This indicates the air density under standard atmospheric conditions. This represents the heat exchange experience index, with a preferred value range of 0.5 to 0.8. Indicates the ventilation or natural wind speed of the transformer substation; Indicates the reference wind speed; This represents the wind speed correction factor.

[0074] Step 2: Based on the basic observation dataset, establish a multi-physics coupled state equation covering condensation formation, electro-thermal coupling temperature rise, and insulation moisture degradation; using historical stable operation samples and historical condensation and discharge fault samples, train and extract discriminative features that reflect the risk of insulation degradation and partial discharge induced by condensation, and generate an initial twin agent and its state space representation that dynamically maps to physical entities.

[0075] In some embodiments, the multiphysics coupled state equations include an electro-thermal coupled temperature rise model and an insulation state degradation model. The electro-thermal coupled temperature rise model is shown in the following equation:

[0076] in, This represents the heat capacity parameter; Indicates the temperature of key components; Indicates time; Indicates power loss; Indicates the heat transfer coefficient; Indicates the heat dissipation area; Indicates ambient temperature; Represents the flow field disturbance term; The insulation degradation model is shown in the following equation:

[0077]

[0078] in, express The level of insulation health at all times; Indicates the initial insulation health; Indicates the temperature rise degradation coefficient; Indicates the cumulative stress due to temperature rise; express Temperature of key components at any given moment; This indicates the reference temperature threshold for insulating materials.

[0079] The temperature rise degradation coefficient It is used to characterize the effect of cumulative stress per unit temperature rise on the rate of decay of insulation health, and its physical meaning is the aging sensitivity of insulation materials under long-term thermal stress. The value of is determined in the following way:

[0080] When historical maintenance data, insulation resistance, dielectric loss factor, partial discharge data, or oil chromatography data are available, the data can be obtained by inverting historical insulation health data; when complete historical degradation data is lacking... Initial values ​​can be given based on the type of insulation material, temperature rise level, manufacturer test data, and operating experience of similar box-type transformers, and can be corrected in subsequent operation through least squares or Bayesian update methods.

[0081] The state space representation of the twin agent is constructed as shown in the following equation:

[0082] in, express The state-space representation of a twin agent at any given moment; Indicates component-level features; Indicates loop-level characteristics; Indicates interval-level features; Indicates the characteristics of an event chain; This represents the state fusion mapping function.

[0083] Step 3: Calculate the confidence interval and model availability index based on the residual between the condensation-related state prediction value output by the twin agent and the actual monitoring quantity. When the prediction residual exceeds the preset boundary, the availability is lower than the set threshold, or the consistency of evidence between electrical quantity, temperature quantity, event record and event chain is lower than the set threshold, the twin agent is determined to be in the untrusted area and self-evolution update is triggered.

[0084] In some embodiments, the calculation of the confidence interval includes: The model predicts residuals and constructs confidence boundaries, as shown in the following equation:

[0085] in, express The model prediction residuals at time step; Indicates the actual monitored value; Indicates the model's predicted value; The credibility boundary is shown in the following formula:

[0086] in, Indicates the credibility boundary; This represents the mean of the residuals; Indicates the standard deviation of the residuals; This represents the confidence coefficient. The confidence coefficient... The width of the confidence boundary is used to control the confidence level, and its value corresponds to the target confidence level; preferably, when the box-type transformer is in a historically stable operation phase, the value is taken as follows: This corresponds to approximately 95% statistical confidence; when in a period of strong disturbance, sudden temperature change, rapid load fluctuation, or high-risk operation, take... This corresponds to a wider range of high-reliability confidence; when the system needs to improve anomaly detection sensitivity, take... When the system needs to reduce the false alarm rate, take .

[0087] The confidence coefficient The dynamic adjustment is based on the residual distribution, as shown in the following formula:

[0088] in, express The dynamic confidence coefficient at time t; Indicates the baseline confidence coefficient; Represents the residual coefficient of variation; This represents the residual skewness or abnormal fluctuation statistic. Indicates the risk level of the event; , and This represents the corresponding weighting coefficient. When residual volatility increases or the event risk level rises, the weighting coefficient is increased. To avoid frequent false triggering of the model under short-term perturbations; when the residuals are stable over a long period and the model output consistency is high, reduce To improve the ability to identify subtle anomalies.

[0089] In some embodiments, the availability metric for the computational model includes: The model availability metric is established as shown in the following formula:

[0090] in, Indicates model availability; This indicates the number of samples that exceed the confidence boundary within the statistical window; This indicates the total number of samples within the statistical window. When... When the availability value is higher than the preset availability threshold, it indicates that the model still has availability under the current operating conditions; when If the value is below a preset threshold, it indicates that the model output lacks credibility and requires an update decision process. Preferably, the availability threshold is set to 0.90 to 0.95.

[0091] The consistency of the model output is analyzed under multiple internal perturbations, as shown in the following equation:

[0092] in, Indicates consistency index; Indicates the number of disturbance tests; Indicates the first Model output under the second perturbation; This represents the average of multiple perturbation outputs; Indicates the perturbation threshold; Indicates characteristic functions; When satisfied Exceeding the preset threshold Below the preset availability threshold or If any condition falls below the preset consistency threshold, the current twin agent is determined to be in an untrusted region and an update determination process is triggered.

[0093] Step 4: Construct the target domain dataset from the untrusted samples and their temporal neighborhood samples, and construct the source domain dataset from the historical stable operation samples, historical condensation and discharge fault samples, and the calibrated trustworthy samples. Use the twin agent that triggers the update moment as the source domain model, inherit the basic temporal feature extraction and event chain encoding capabilities of the source domain model through transfer learning, and reduce the distribution difference between the source domain model and the target domain in the fusion feature space by adopting the domain adversarial feature adaptation method. Freeze the bottom feature extraction layer, fine-tune the state fusion layer and the output decision layer, and apply step size constraints, convergence criteria and multi-source evidence consistency constraints to make the model converge to the new trustworthy state boundary and complete the self-evolution update.

[0094] In some embodiments, the source domain dataset is as follows:

[0095] in, Represents the source domain dataset. This indicates the multi-source monitoring characteristics under historical stable operating conditions and historical fault conditions. This indicates the corresponding status label or health assessment result. Indicates the characteristics of a historical event chain; The target domain dataset is shown in the following formula:

[0096] in, Represents the target domain dataset. This represents the multi-source monitoring characteristics of the current untrusted sample and the time windows before and after it. The target domain event chain features are represented. Target domain samples may not have artificial labels; they can participate in model updates through trusted pseudo-labels, multi-source consistency constraints, or subsequent confirmed labels. Model updates are performed using a combination of parameter transfer and domain adversarial feature adaptation. The twin agent model includes a bottom-level general feature extraction layer, a middle-level multi-source state fusion layer, an environment adaptation layer, and an output decision layer. The bottom-level general feature extraction layer extracts basic temporal features of electrical quantities, temperature quantities, and event chains. The middle-level multi-source state fusion layer establishes the coupling relationship between electrical quantities, temperature quantities, insulation, and events. The environment adaptation layer characterizes the distribution shift caused by high-altitude and cold environments. The output decision layer outputs the state evaluation results. During the transfer process, the parameters of the bottom-level general feature extraction layer are frozen, and the parameters of the source domain's trained electrical feature encoding layer, temperature feature encoding layer, and event chain encoding layer are transferred. The middle-level multi-source state fusion layer and output decision layer are fine-tuned with small steps. When the target domain samples continuously meet the consistency requirements and the number reaches a set scale, the environment adaptation layer is unfrozen and updated. The difference in feature distribution between the source domain and the target domain is constrained by domain adversarial loss or maximum mean difference loss.

[0097] In some embodiments, the self-evolutionary update involves updating the total loss function of the model. Build as follows:

[0098] in, For prediction error loss; This represents a loss due to multi-source consistency. For regularization terms; Losses due to domain confrontation; The prediction error loss Calculate using the following formula:

[0099] in, This indicates the number of samples that participated in this update; Indicates the first The actual label, credible pseudo label, or historical confirmation status of each sample; The model represents the first The predicted output for each sample; The multi-source consistency loss Calculate using the following formula:

[0100] in, and These represent inputs from different data sources; and These represent the feature mapping functions of the corresponding data sources; Indicates data source With data source Consistency weights are used to constrain the consistency of state evidence among electrical quantities, temperature quantities, SOE events, and station-level event chains; The regularization term Calculate using the following formula:

[0101] in, This represents the currently updated model parameters; This represents the parameters of the source domain model before the update is triggered; this regularization term is used to limit the new model from deviating too far from the historical reliable model, thereby maintaining the original normal operating condition recognition capability.

[0102] The domain adversarial loss Calculate using the following formula:

[0103] in, Indicates a feature extractor; Representation domain discriminator; Indicates the first One source domain sample; Indicates the first One target domain sample; and These represent the number of samples in the source domain and the target domain, respectively.

[0104] In some embodiments, during the self-evolutionary update process, step size constraints are applied to the model parameters and a convergence criterion is established; Model parameter update increments are truncated using the following formula:

[0105] Model parameters are updated using the following formula:

[0106] in, Indicates the first Model parameters at the next update; Indicates the first Parameter increment during the next update; Indicates the learning rate; This represents the gradient of the total loss function with respect to the model parameters; Indicates the upper limit of the step size constraint; This indicates that the gradient update amount is limited to A cutoff function within the range, when the gradient update exceeds At that time, limit it to Within a certain range, thus preventing unreliable or noisy samples from causing drastic changes in model parameters; The upper limit of the step size constraint The target domain sample consistency is adaptively determined, as shown in the following formula:

[0107] in, Indicates the upper limit of the baseline step size; This represents the output consistency index of the target domain samples under multiple internal perturbations. This represents the adjustment coefficient; when the consistency of the target domain samples is high, the update step size can be moderately increased; when the consistency of the target domain samples is low, the update step size should be decreased to ensure update stability. The convergence criterion is shown in the following formula:

[0108] in, and They represent the first Next and first The loss value of the next iteration; This represents the convergence threshold. When the convergence criterion is met for multiple consecutive update cycles, and the updated model meets the requirements of residual boundary, consistency index, and model availability within the verification window, the twin agent is deemed to have completed a reliable self-evolutionary update.

[0109] Step 5: Utilize the updated twin agent to output insulation health and condensation-triggered discharge risk values ​​in real time. Based on the condensation-triggered discharge risk value and current environmental parameters, generate dehumidification linkage control commands to control the start / stop and power adjustment of the heater or dehumidifier, thereby achieving closed-loop linkage between condensation protection and discharge risk suppression.

[0110] In some embodiments, the updated trusted twin agent is used to output insulation health in real time. Risk value of discharge triggered by condensation ,in The twin agent has a built-in three-level threshold: the first threshold... Second threshold Danger threshold .when When the dehumidifier and heater are in standby or low-power monitoring mode, only basic air circulation within the cabinet is maintained to avoid unnecessary energy consumption; when When a dehumidification start command is generated, the dehumidifier operates at a preset base power, reducing the absolute humidity inside the cabinet through condensation or rotary dehumidification, while continuously monitoring the temperature difference between the cabinet interior and the dew point temperature; when At that time, while the dehumidifier is running continuously, a heater start command is generated, and the real-time dew point temperature is superimposed. Using the safe superheat level as the target temperature, a proportional-integral control algorithm is employed to adjust the heater power, ensuring that the temperature in the critical insulation area inside the cabinet remains consistently above the dew point temperature, actively disrupting condensation formation conditions and accelerating the drying of the insulation surface; when At this time, both the dehumidifier and heater are forced to switch to maximum rated power operation, and high-priority alarm signals and event logs are output to alert operators. During the dehumidification linkage control process, the twin agent continuously receives feedback on cabinet temperature and humidity, load current, and partial discharge monitoring data after heating and dehumidification actions, updates insulation health and risk values ​​in real time, and dynamically adjusts control outputs until... Falling back to The following constitutes a closed-loop linkage between condensation protection and discharge risk suppression.

[0111] Based on the same inventive concept, and corresponding to any of the above embodiments, refer to... Figure 2 This invention provides a risk identification and dehumidification linkage control system for condensation-triggered discharge of box-type transformers in high-altitude and cold regions, used to implement the aforementioned method for risk identification and dehumidification linkage control of condensation-triggered discharge of box-type transformers in high-altitude and cold regions, including: The multi-source data synchronization and compensation module is used to acquire multi-source heterogeneous monitoring data during the operation of the box-type transformer. It eliminates clock deviation by aligning with a unified time base, and performs joint compensation for data loss, sensor drift and communication delay anomalies, and outputs a basic observation dataset with consistent timing. The initialization module, connected to the multi-source data synchronization and compensation module, is used to receive the basic observation dataset, establish a multi-physics field coupled state equation covering condensation formation, electro-thermal coupling temperature rise and insulation moisture degradation, train and extract discriminative features using historical stable operation samples and historical condensation and discharge fault samples, and generate an initial twin intelligent agent and its state space expression that is dynamically mapped to the physical entity. The self-evolution trigger module, connected to the initialization module, is used to calculate the confidence interval and model availability index based on the residual between the condensation-related state prediction value output by the twin agent and the actual monitoring quantity, and to evaluate the consistency of multi-source evidence among electrical quantities, temperature quantities, event records and event chains. When the prediction residual exceeds the preset boundary, the availability is lower than the set threshold, or the evidence consistency is lower than the set threshold, the twin agent is determined to be in an untrustworthy area and a self-evolution trigger signal is issued. The domain adaptive update module, connected to the self-evolution trigger module, receives the self-evolution trigger signal, constructs the target domain dataset from samples deemed untrustworthy and their temporal neighborhood samples, and constructs the source domain dataset from historical stable operation samples, historical condensation and discharge fault samples, and calibrated trustworthy samples. The twin agent at the trigger update moment is used as the source domain model. Through transfer learning, it inherits the basic temporal feature extraction and event chain encoding capabilities of the source domain model. It adopts a domain adversarial feature adaptation method to reduce the distribution difference between the source domain model and the target domain in the fusion feature space, freezes the bottom feature extraction layer, fine-tunes the state fusion layer and the output decision layer, and applies step size constraints, convergence criteria and multi-source evidence consistency constraints to make the model converge to the new trustworthy state boundary, complete the self-evolution update, and output the updated twin agent parameters. The risk identification module, connected to the domain adaptive update module, is used to load the updated twin agent parameters, calculate and output the insulation health and condensation-triggered discharge risk values ​​in real time. The dehumidification linkage control module, connected to the risk identification module, is used to receive the condensation-triggered discharge risk value and current environmental parameters, generate dehumidification linkage control commands, and control the start / stop and power adjustment of the heater or dehumidifier to achieve closed-loop linkage between condensation protection and discharge risk suppression.

[0112] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the method for identifying and dehumidifying the risk of condensation-triggered discharge in a high-altitude and cold-climate box-type transformer according to the embodiment.

[0113] Alternatively, the aforementioned electronic device may be a server.

[0114] In addition, this embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for identifying the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers and for dehumidification linkage control.

[0115] It is understood that the processor in the embodiments of the present invention may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0116] The method steps in the embodiments of the present invention can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0117] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a storage medium or transmitted through a storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive (SSD)).

Claims

1. A method for identifying the risk of condensation-triggered discharge and linking it with dehumidification control in a high-altitude, cold-climate box-type transformer, characterized in that... Includes the following steps: Acquire multi-source heterogeneous monitoring data during the operation of box-type transformers, eliminate clock deviations between different data sources by aligning the multi-source heterogeneous monitoring data with a unified time reference, jointly compensate for data loss, sensor drift and communication delay anomalies, and construct a time-consistent basic observation dataset. Based on the basic observation dataset, a multi-physics field coupled state equation covering condensation formation, electro-thermal coupling temperature rise, and insulation moisture degradation is established. Using historical stable operation samples and historical condensation and discharge fault samples, discriminative features reflecting the risk of insulation degradation and partial discharge induced by condensation are trained and extracted to generate an initial twin intelligent agent and its state space expression that is dynamically mapped to physical entities. Based on the residual between the condensation-related state prediction value output by the twin agent and the actual monitoring quantity, the confidence interval and model availability index are calculated. When the prediction residual exceeds the preset boundary, the availability is lower than the set threshold, or the evidence consistency between electrical quantity, temperature quantity, event record and event chain is lower than the set threshold, the twin agent is determined to be in the untrusted area and self-evolution update is triggered. The untrusted samples and their temporal neighborhood samples are constructed as the target domain dataset, while historical stable operation samples, historical condensation and discharge fault samples, and calibrated trustworthy samples are constructed as the source domain dataset. The twin agent that triggers the update moment is used as the source domain model. The basic temporal feature extraction and event chain encoding capabilities of the source domain model are inherited through transfer learning. The domain adversarial feature adaptation method is used to reduce the distribution difference between the source domain model and the target domain in the fusion feature space. The bottom feature extraction layer is frozen, the state fusion layer and the output decision layer are fine-tuned, and step size constraints, convergence criteria and multi-source evidence consistency constraints are applied to make the model converge to the new trustworthy state boundary and complete the self-evolution update. The updated twin agent outputs insulation health and condensation-triggered discharge risk values ​​in real time. Based on the condensation-triggered discharge risk value and current environmental parameters, it generates dehumidification linkage control commands to control the start-stop and power adjustment of the heater or dehumidifier, realizing closed-loop linkage between condensation protection and discharge risk suppression.

2. The method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers as described in claim 1, characterized in that, The unified time reference alignment is achieved by constructing a weighted fused time reference sequence, as shown in the following equation: in, Represents a unified time reference series; Represents a Network Time Protocol (NTP) timestamp; Indicates a Precision Time Protocol (PTP) timestamp; Indicates satellite time synchronization timestamp; , and Let represent the weight coefficients of the corresponding time synchronization sources, and satisfy . , , , ; The weighting coefficients are determined by the standard deviation of the time deviation of each time source, as shown in the following formula: in, Indicates the first The fusion weights of time-series sources; Indicates the first The standard deviation of the time deviation of the time source within the statistical window; time window resampling of data with different sampling frequencies based on a unified time reference sequence, as shown in the following formula: in, This represents the data after resampling; Indicates the first time within the time window The original data of each sampling point; This indicates the number of sampling points within the time window.

3. The method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers as described in claim 1, characterized in that... The joint compensation for data loss, sensor drift, and communication delay anomalies includes: For short-term missing data, an interpolation reconstruction method is used for recovery, as shown in the following formula: in, Indicates missing moments The recovery value; and These represent the observations at both ends of the missing interval; For continuous states, Kalman smoothing is used for correction, and the state update expression is: in, This represents the state estimate at the previous moment; This represents the estimated prior state value at the current moment; This represents the estimated posterior state at the current time. Represents the state transition matrix; Represents the control input matrix; Indicates external control input; Represents the prior error covariance matrix; Represents the posterior error covariance matrix; Represents the process noise covariance matrix; Indicates Kalman gain; Represents the observed value; Represents the observation matrix; Represents the observation noise covariance matrix; Represents the identity matrix; The state vector is set as shown in the following formula: in, Indicates voltage status; Indicates the current state; Indicates power status; This indicates the oil temperature or the equivalent temperature of the windings. Indicates the ambient temperature status; Indicates the insulation health status; This indicates the event chain state code.

4. The method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers as described in claim 3, characterized in that, In high-altitude and cold environments, the process noise covariance matrix and the observation noise covariance matrix are dynamically corrected as shown in the following equation: in, This represents the dynamically corrected process noise covariance matrix; Represents the reference process noise covariance matrix; This represents the dynamically corrected observation noise covariance matrix; Represents the reference observation noise covariance matrix; Indicates the change in ambient temperature; Indicates the change in altitude; Indicates the intensity of sensor fluctuations; This represents the communication delay statistics; , , and This is the corresponding empirical correction factor; A high-altitude environment correction factor is introduced to correct the heat transfer coefficient, as shown in the following formula: in, This represents the corrected heat transfer coefficient; This represents the baseline heat transfer coefficient under standard altitude conditions. This indicates the correction factor for high-altitude environments; Indicates altitude and ambient temperature air density below; This indicates the air density under standard atmospheric conditions. Indicates the heat exchange experience index; Indicates the ventilation or natural wind speed of the transformer substation; Indicates the reference wind speed; This represents the wind speed correction factor.

5. The method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers as described in claim 1, characterized in that... The multiphysics coupled state equations include an electro-thermal coupled temperature rise model and an insulation state degradation model. The electro-thermal coupled temperature rise model is shown in the following equation: in, This represents the heat capacity parameter; Indicates the temperature of key components; Indicates time; Indicates power loss; Indicates the heat transfer coefficient; Indicates the heat dissipation area; Indicates ambient temperature; Represents the flow field disturbance term; The insulation degradation model is shown in the following equation: in, express The level of insulation health at all times; Indicates the initial insulation health; Indicates the temperature rise degradation coefficient; Indicates the cumulative stress due to temperature rise; express Temperature of key components at any given moment; Indicates the reference temperature threshold for insulating materials; The state space representation of the twin agent is constructed as shown in the following equation: in, express The state-space representation of a twin agent at any given moment; Indicates component-level features; Indicates loop-level characteristics; Indicates interval-level features; Indicates the characteristics of an event chain; This represents the state fusion mapping function.

6. The method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers as described in claim 1, characterized in that... The confidence interval for calculation includes: The model predicts residuals and constructs confidence boundaries, as shown in the following equation: in, express The model prediction residuals at time step; This represents the actual monitored value; Indicates the model's predicted value; The credibility boundary is shown in the following formula: in, Indicates the credibility boundary; This represents the mean of the residuals; Indicates the standard deviation of the residuals; Indicates the confidence coefficient; The confidence coefficient The dynamic adjustment is based on the residual distribution, as shown in the following formula: in, express The dynamic confidence coefficient at time t; Indicates the baseline confidence coefficient; Represents the residual coefficient of variation; This represents the residual skewness or abnormal fluctuation statistic. Indicates the risk level of the event; , and This represents the corresponding weighting coefficient.

7. The method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers as described in claim 6, characterized in that... The availability metrics of the computational model include: The model availability metric is established as shown in the following formula: in, Indicates model availability; This indicates the number of samples that exceed the confidence boundary within the statistical window; This indicates the total number of samples within the statistics window; The consistency of the model output is analyzed under multiple internal perturbations, as shown in the following equation: in, Indicates consistency index; Indicates the number of disturbance tests; Indicates the first Model output under the second perturbation; This represents the average of multiple perturbation outputs; Indicates the perturbation threshold; Indicates characteristic functions; When satisfied Exceeding the preset threshold Below the preset availability threshold or If any condition falls below the preset consistency threshold, the current twin agent is determined to be in an untrusted region and an update determination process is triggered.

8. The method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers as described in claim 1, characterized in that, In the self-evolutionary update, the model updates the total loss function. Build as follows: in, For prediction error loss; This represents a loss due to multi-source consistency. For regularization terms; Losses due to domain confrontation; The prediction error loss Calculate using the following formula: in, This indicates the number of samples that participated in this update; Indicates the first The actual label, credible pseudo label, or historical confirmation status of each sample; The model represents the first The predicted output for each sample; The multi-source consistency loss Calculate using the following formula: in, and These represent inputs from different data sources; and These represent the feature mapping functions of the corresponding data sources; Indicates data source With data source Consistency weights are used to constrain the consistency of state evidence among electrical quantities, temperature quantities, SOE events, and station-level event chains; The regularization term Calculate using the following formula: in, This represents the currently updated model parameters; This represents the parameters of the source domain model before the update was triggered; The domain adversarial loss Calculate using the following formula: in, Indicates a feature extractor; Representation domain discriminator; Indicates the first One source domain sample; Indicates the first One target domain sample; and These represent the number of samples in the source domain and the target domain, respectively.

9. The method for identifying and controlling the risk of condensation-triggered discharge in high-altitude and cold-climate box-type transformers as described in claim 1, characterized in that, During the self-evolutionary update process, step size constraints are applied to the model parameters and convergence criteria are established. Model parameter update increments are truncated using the following formula: Model parameters are updated using the following formula: in, Indicates the first Model parameters at the next update; Indicates the first Parameter increment during the next update; Indicates the learning rate; This represents the gradient of the total loss function with respect to the model parameters; Indicates the upper limit of the step size constraint; This indicates that the gradient update amount is limited to A cutoff function within a specified range; The upper limit of the step size constraint The target domain sample consistency is adaptively determined, as shown in the following formula: in, Indicates the upper limit of the baseline step size; This represents the output consistency index of the target domain samples under multiple internal perturbations. Indicates the adjustment coefficient; The convergence criterion is shown in the following formula: in, and They represent the first Next and first The loss value of the next iteration; This represents the convergence threshold. When the convergence criterion is met for multiple consecutive update cycles, and the updated model meets the requirements of residual boundary, consistency index, and model availability within the verification window, the twin agent is deemed to have completed a reliable self-evolutionary update.

10. A risk identification and dehumidification linkage control system for condensation-triggered discharge of a high-altitude, cold-climate box-type transformer, used to implement the risk identification and dehumidification linkage control method for condensation-triggered discharge of a high-altitude, cold-climate box-type transformer as described in any one of claims 1-9, characterized in that... include: The multi-source data synchronization and compensation module is used to acquire multi-source heterogeneous monitoring data during the operation of the box-type transformer. It eliminates clock deviation by aligning with a unified time base, and performs joint compensation for data loss, sensor drift and communication delay anomalies, and outputs a basic observation dataset with consistent timing. The initialization module, connected to the multi-source data synchronization and compensation module, is used to receive the basic observation dataset, establish a multi-physics field coupled state equation covering condensation formation, electro-thermal coupling temperature rise and insulation moisture degradation, train and extract discriminative features using historical stable operation samples and historical condensation and discharge fault samples, and generate an initial twin intelligent agent and its state space expression that is dynamically mapped to the physical entity. The self-evolution trigger module, connected to the initialization module, is used to calculate the confidence interval and model availability index based on the residual between the condensation-related state prediction value output by the twin agent and the actual monitoring quantity, and to evaluate the consistency of multi-source evidence among electrical quantities, temperature quantities, event records and event chains. When the prediction residual exceeds the preset boundary, the availability is lower than the set threshold, or the evidence consistency is lower than the set threshold, the twin agent is determined to be in an untrustworthy area and a self-evolution trigger signal is issued. The domain adaptive update module, connected to the self-evolution trigger module, receives the self-evolution trigger signal, constructs the target domain dataset from samples deemed untrustworthy and their temporal neighborhood samples, and constructs the source domain dataset from historical stable operation samples, historical condensation and discharge fault samples, and calibrated trustworthy samples. The twin agent at the trigger update moment is used as the source domain model. Through transfer learning, it inherits the basic temporal feature extraction and event chain encoding capabilities of the source domain model. It adopts a domain adversarial feature adaptation method to reduce the distribution difference between the source domain model and the target domain in the fusion feature space, freezes the bottom feature extraction layer, fine-tunes the state fusion layer and the output decision layer, and applies step size constraints, convergence criteria and multi-source evidence consistency constraints to make the model converge to the new trustworthy state boundary, complete the self-evolution update, and output the updated twin agent parameters. The risk identification module, connected to the domain adaptive update module, is used to load the updated twin agent parameters, calculate and output the insulation health and condensation-triggered discharge risk values ​​in real time. The dehumidification linkage control module, connected to the risk identification module, is used to receive the condensation-triggered discharge risk value and current environmental parameters, generate dehumidification linkage control commands, and control the start / stop and power adjustment of the heater or dehumidifier to achieve closed-loop linkage between condensation protection and discharge risk suppression.

Images