A transformer real-time data abnormality early warning method and system

Abstract

The present application relates to the technical fields of data processing, anomaly detection and computer program control, in particular to a transformer real-time data anomaly early warning method and system. The method obtains real-time data of voltage, current, load rate, active power, reactive power, oil temperature, winding temperature and environmental temperature in the operation process of the transformer, forms a multivariate operation state sequence after time stamp alignment, missing compensation and spike suppression; constructs electrical excitation factors, heat dissipation boundary factors and thermal inertia response factors according to the sequence, inputs a mechanism constraint time series reconstruction model, and generates an expected state sequence under a normal electric-thermal coupling state; and calculates a residual sequence between the actual sequence and the expected sequence. The present application can distinguish between normal thermal responses caused by load fluctuations and environmental changes and abnormal deviation states, and improve the stability and interpretability of transformer real-time data anomaly early warning.

Description

Technical Field

[0001] This invention relates to the fields of data processing, anomaly detection and computer program control technology, specifically to a method and system for real-time data anomaly early warning of transformers. Background Technology

[0002] Real-time data anomaly early warning for transformers typically relies on online monitoring platforms or station-end acquisition devices to obtain operational data. Commonly used data include voltage, current, load rate, active power, reactive power, oil temperature, winding temperature, and ambient temperature. Existing solutions generally upload data according to a fixed sampling period, and perform time sorting, missing value imputation, limit violation judgment, and trend display in the background system. For anomaly identification, a common method is to set upper limit values, rate of change thresholds, or duration thresholds for parameters such as oil temperature, winding temperature, current, and load rate. When a monitored quantity exceeds the preset threshold, an alarm is triggered. Some solutions combine historical averages, moving averages, or experience ranges of similar equipment to adjust the thresholds, allowing the alarm boundaries to be adjusted to a limited extent with seasonality or load levels. This type of solution is simple to implement and easy to deploy in engineering, but it essentially still uses the value exceeding the limit of a single variable or a small number of variables as the judgment basis, and the electrothermal response relationship between variables is not included in the core judgment process.

[0003] In existing data analysis-based early warning schemes, the conventional approach is to input real-time transformer data into time series prediction models, clustering models, or anomaly detection models. The models then learn the distribution characteristics of historical operating data and provide predicted values ​​or anomaly scores for subsequent real-time data. These schemes typically first standardize data such as voltage, current, load rate, oil temperature, and winding temperature, then organize them into sample sequences according to fixed time windows. Neural networks, regression models, distance metric models, or probabilistic models are then used to calculate the deviation of the current data from historical samples. To reduce the impact of occasional fluctuations, the system may smooth the anomaly scores or require anomalies to occur within multiple consecutive sampling periods before triggering an early warning. While this type of scheme can utilize a large amount of historical data, the training results are primarily affected by the sample distribution. When there are insufficient fault samples, significant changes in operating conditions, or incomplete historical data coverage, the model may easily identify changes in normal operating conditions as anomalies or absorb early anomalies as normal fluctuations.

[0004] Transformer operation exhibits strong electrothermal coupling characteristics. Changes in electrical quantities such as current, load factor, active power, and reactive power typically cause variations in oil and winding temperatures. However, the temperature response is not instantaneous but is influenced by ambient temperature, heat dissipation conditions, heat capacity, and operating duration, manifesting as a delayed rise, slow accumulation, or delayed fall. Some existing solutions introduce a correlation between load factor and oil temperature or correct temperature thresholds based on ambient temperature. However, these solutions typically only incorporate environmental factors as compensation items in threshold calculations, failing to establish a joint modeling of electrical excitation factors, heat dissipation boundary factors, and thermal inertia response factors. Consequently, under conditions such as sudden load increases, short-term load decreases, and continuous changes in ambient temperature, the system struggles to determine whether temperature changes conform to a normal thermal inertia response. Anomaly warnings still rely on local numerical judgments, lacking constraints on the multi-variable coordinated deviation process.

[0005] The lack of a unified data processing mechanism capable of simultaneously expressing instantaneous changes in electrical quantities, hysteretic responses to temperature, changes in heat dissipation boundaries, and multivariate residual coupling relationships makes it difficult to distinguish between normal electrothermal coupling states and abnormal deviation states when load fluctuations, ambient temperature changes, and normal thermal inertia responses coexist. Fixed threshold methods cannot reflect the reasonable temperature rise process under different operating conditions, while purely data-driven models are easily affected by historical sample distributions and deviate from the transformer's operating mechanism. When there is a time difference between changes in current or load rate and changes in oil temperature and winding temperature, existing solutions tend to judge normal hysteresis responses as abnormal; when the temperature residual continues to accumulate after the load decreases, or when temperature changes do not match electrical excitation, existing solutions may delay warnings because parameters have not exceeded limits. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for real-time data anomaly early warning of transformers, which can effectively solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for real-time data anomaly early warning of transformers includes: acquiring real-time data of voltage, current, load rate, active power, reactive power, oil temperature, winding temperature and ambient temperature during transformer operation; aligning, compensating for missing data and suppressing spikes according to the sampling timestamps to obtain a multivariable operating state sequence. Based on the multivariate operating state sequence, an electrical excitation factor, a heat dissipation boundary factor, and a thermal inertia response factor are constructed. The electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor are then input into a mechanism-constrained time-series reconstruction model to generate the desired state sequence under normal electrothermal coupling conditions. Calculate the residual sequence between the multivariate running state sequence and the desired state sequence, and generate an anomaly confidence index based on residual magnitude, residual duration, residual direction consistency and intervariate residual coupling degree; Based on the adaptive early warning boundary formed by the historical normal operation range, the abnormal confidence index is classified and judged, and the real-time data abnormality early warning result of the transformer is output.

[0008] Preferably, constructing the electrical excitation factor, heat dissipation boundary factor, and thermal inertia response factor based on the multivariate operating state sequence includes: extracting the load change amplitude, load change slope, and power component change relationship from the current, the load rate, the active power, and the reactive power to form the electrical excitation factor; The heat dissipation temperature difference, environmental change trend, and temperature baseline shift are extracted from the ambient temperature, the oil temperature, and the winding temperature to form the heat dissipation boundary factor. The thermal inertia response factor is formed based on the hysteretic changes in the oil temperature and the winding temperature relative to the electrical excitation factor. The electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor are combined into an electrothermal coupling feature sequence according to the same time index.

[0009] Preferably, inputting the electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor into the mechanism constraint time series reconstruction model includes: dividing the electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor into a short-term impact window, a medium-term load fluctuation window, and a long-term temperature rise accumulation window; The variable change trajectories in different time windows are encoded to obtain multi-scale working condition representations; In the mechanism-constrained time-series reconstruction model, constraint terms are set for the lag changes in electrical quantities and temperature quantities. The multi-scale operating condition characterization and the constraint terms are input into the reconstruction calculation process to obtain the desired state sequence corresponding to the current operating condition.

[0010] Preferably, the step of generating anomaly confidence indices based on residual amplitude, residual duration, residual direction consistency, and intervariate residual coupling degree includes: calculating the standardized deviation values ​​of electrical quantity residuals, temperature quantity residuals, and environmental quantity residuals, respectively; The duration of the residual and the consistency of the residual direction are determined based on whether the residuals of the same variable maintain the same direction of deviation within consecutive sampling periods. The degree of coupling of the intervariable residuals is determined based on the matching relationship between the electrical quantity residuals and the temperature quantity residuals within a preset hysteresis window. The anomaly confidence index is obtained by weighting and fusing the standardized deviation, the residual duration, the residual directional consistency, and the intervariate residual coupling degree.

[0011] Preferably, forming the thermal inertia response factor includes: using the change time of the electrical excitation factor as a reference, determining the candidate hysteresis intervals where the oil temperature and the winding temperature show corresponding changes; Calculate the temperature rise slope, cumulative temperature rise, and temperature rise fall slope within the candidate hysteresis interval; The temperature rise slope, the cumulative temperature rise, and the temperature rise fall slope are jointly normalized with the heat dissipation boundary factor to obtain the thermal response state vector. The thermal response state vector within a continuous sampling period is paired with the electrical excitation factor to form the thermal inertial response factor, which includes the load increase response, the load hold response, and the load decrease response.

[0012] Preferably, in the mechanism-constrained time-series reconstruction model, a constraint term for the hysteretic change of electrical quantity and the hysteretic change of temperature quantity is set, including: generating hysteretic response templates under different load ranges based on the correspondence between the electrical excitation factor and the thermal inertia response factor in the historical normal operation range; The multi-scale working condition representation input in real time is matched with the hysteresis response template to obtain the response constraint weights; When reconstructing the expected values ​​of the oil temperature and the winding temperature, the response constraint weights are embedded in the reconstruction loss calculation, so that the expected state sequence simultaneously includes constraints on instantaneous changes in electrical quantities and constraints on delayed changes in temperature quantities.

[0013] Preferably, determining the intervariable residual coupling degree based on the matching relationship between the electrical quantity residual and the temperature quantity residual within a preset lag window includes: taking the abrupt change point of the electrical quantity residual as the starting point of the window, and extracting the temperature quantity residual subsequence under different lag step sizes; Calculate the proportion of direction alignment, slope correspondence, and cumulative deviation between the electrical quantity residual and each temperature quantity residual subsequence; The residual coupling vector is determined based on the direction alignment ratio, the slope correspondence ratio, and the cumulative deviation ratio. The intervariable residual coupling degree is obtained based on the distance between the residual coupling vector and the historical normal coupling vector.

[0014] Preferably, pairing the thermal response state vector with the electrical excitation factor within a continuous sampling period includes: dividing the electrical excitation factor into segments according to the state transition sequence of load increase, load hold, and load decrease; Extract the initial value, peak value, fall value, and hysteresis length of the thermal response state vector within each segment; The initial value, peak value, fallback value and hysteresis length between adjacent segments are sequentially concatenated to form a response memory matrix; The response memory matrix is ​​used as an additional input to the mechanism-constrained temporal reconstruction model, and together with the electrothermal coupling feature sequence, it participates in the generation of the desired state sequence.

[0015] Preferably, the adaptive early warning boundary formed based on the historical normal operation range includes: dividing the boundary sample clusters within the historical normal operation range according to the load range and the ambient temperature range; Calculate the quantile boundaries of the anomaly confidence index and the stability boundaries of the intervariate residual coupling degree in each boundary sample cluster; The load range and ambient temperature range corresponding to the real-time data are matched to the target boundary sample cluster, and interpolation is performed according to the boundary difference of adjacent boundary sample clusters to obtain the adaptive warning boundary under the current working condition. The anomaly confidence index is compared with the adaptive early warning boundary to generate the real-time data anomaly early warning result for the transformer.

[0016] A real-time data anomaly early warning system for transformers includes: a data processing module, used to acquire real-time data of voltage, current, load rate, active power, reactive power, oil temperature, winding temperature and ambient temperature during transformer operation, and to perform alignment, missing data compensation and spike suppression according to the sampling timestamp to obtain a multivariable operating state sequence; An electrothermal coupling feature construction module is used to construct electrical excitation factors, heat dissipation boundary factors, and thermal inertia response factors based on the multivariate operating state sequence. The mechanism-constrained residual reconstruction module is used to generate the desired state sequence under normal electrothermal coupling state based on the electrical excitation factor, the heat dissipation boundary factor and the thermal inertia response factor, and to calculate the residual sequence between the multivariate operating state sequence and the desired state sequence. The anomaly confidence calculation module is used to generate anomaly confidence indices based on residual magnitude, residual duration, residual direction consistency, and intervariate residual coupling degree. The early warning determination module is used to classify and determine the abnormal confidence index based on the adaptive early warning boundary formed by the historical normal operation range and output the real-time data abnormality early warning result of the transformer.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By acquiring real-time data on voltage, current, load rate, active power, reactive power, oil temperature, winding temperature, and ambient temperature, and processing this real-time data into a multivariate operating state sequence, and then constructing electrical excitation factors, heat dissipation boundary factors, and thermal inertia response factors, the judgment object for anomaly warning can be transformed from a single monitored quantity to an electrothermal coupling state. The mechanism-constrained time-series reconstruction model generates the expected state sequence under normal electrothermal coupling state based on the current operating conditions, and then uses the residual sequence between the actual sequence and the expected sequence for anomaly judgment, ensuring that changes in electrical quantities, delayed changes in temperature quantities, and changes in environmental boundaries are all within the same calculation link. The residual amplitude, residual duration, residual direction consistency, and intervariate residual coupling degree together form an anomaly confidence index, which can avoid triggering warnings solely due to single-point limit exceedances or short-term fluctuations, and can also identify abnormal deviation states such as continuous accumulation of temperature residuals after load reduction and mismatch between temperature rise slope and electrical excitation, thereby forming a corresponding data processing and judgment mechanism for the above-mentioned main technical problems.

[0018] 2. By forming adaptive early warning boundaries based on load and ambient temperature ranges, and combining short-term impact windows, medium-term load fluctuation windows, and long-term temperature rise accumulation windows for reconstruction calculations, early warning judgments can select corresponding boundaries according to the current operating conditions, reducing the dependence of fixed thresholds on seasonal, load level, and heat dissipation condition changes. The thermal inertia response factor incorporates load increase response, load hold response, and load decrease response into the same state description. The response memory matrix further retains the temperature rise peak, fall value, and lag length between adjacent operating segments, enabling the desired state sequence to reflect the thermal response trajectory during continuous operating condition transitions. The intervariable residual coupling degree is calculated by the proportion of direction consistency, slope correspondence, and cumulative deviation of electrical quantity residuals and temperature quantity residuals within the lag window, enabling abnormal early warning results to distinguish isolated spikes, normal lags, and continuous coupling deviations, improving the stability and interpretability of real-time data anomaly judgment. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the transformer real-time data anomaly early warning method of the present invention; Figure 2 This is a flowchart illustrating the construction of the electrothermal coupling feature sequence of the present invention. Figure 3 This is a flowchart of the mechanism-constrained temporal reconstruction and anomaly confidence index calculation of the present invention; Figure 4 This is a flowchart of the transformer real-time data anomaly early warning system and adaptive early warning boundary processing of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please refer to Figure 1 This embodiment provides a method and system for real-time data anomaly early warning of transformers, including acquiring real-time data of voltage, current, load rate, active power, reactive power, oil temperature, winding temperature, and ambient temperature during transformer operation; aligning, compensating for missing values, and suppressing spikes in the real-time data according to the sampling timestamp to obtain a multivariate operating state sequence; constructing an electrical excitation factor, a heat dissipation boundary factor, and a thermal inertia response factor based on the multivariate operating state sequence; inputting the electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor into a mechanism-constrained time-series reconstruction model to generate a desired state sequence under normal electrothermal coupling; calculating the residual sequence between the multivariate operating state sequence and the desired state sequence; and generating an anomaly confidence index based on the residual amplitude, residual duration, residual direction consistency, and intervariate residual coupling degree; classifying the anomaly confidence index based on an adaptive early warning boundary formed by historical normal operating intervals; and outputting a real-time data anomaly early warning result for the transformer. In specific implementation... The real-time data forms an original matrix according to a unified sampling time. Each row corresponds to a sampling time, and each column corresponds to a type of operating variable. When the sampling times of different variables are inconsistent, data within adjacent time points are selected based on the target sampling time for linear interpolation or maintenance compensation. When a variable is missing in a continuous sampling period, limited compensation is performed based on the changing trend of adjacent variables within the same operating interval and the historical average value under the same operating conditions. Instantaneous spikes exceeding the range of change in adjacent windows are clipped and replaced to ensure that the data entering the subsequent model maintains time consistency and variable correspondence. The working principle of the method is to put the electrical excitation, heat dissipation boundary, and thermal inertia response in transformer operation into the same time-series calculation link, generate the desired state sequence using the electrothermal coupling relationship under normal conditions, and then use the residual between the actual state and the desired state as the basis for anomaly judgment, avoiding early warning based solely on the over-limit of a single monitoring quantity. In this embodiment, it is possible to distinguish between normal thermal response and abnormal deviation when load fluctuations and ambient temperature changes coexist.

[0022] In this embodiment, the multivariate operating state sequence uses a time window as the basic processing object. The window length is determined based on the sampling density of the transformer operating data and the temperature response period. The window contains synchronized data of voltage, current, load factor, active power, reactive power, oil temperature, winding temperature, and ambient temperature. After data alignment, three types of descriptive quantities are established for each variable: operating value, incremental value, and change slope. The operating value is used to retain the current state, the incremental value is used to describe the direction of change between adjacent sampling times, and the change slope is used to describe the intensity of change per unit time. Missing value compensation does not directly use the global mean for replacement, but rather determines the reasonable range of missing variables based on the correlation between load factor, current, and active power within the same time window. For example, if the current is missing but the load factor is missing... When the load factor and active power change continuously, the direction of the current compensation value is constrained by the direction of the load factor change, and the upper and lower bounds of the compensation value are constrained by the amplitude of the active power change. Peak suppression does not delete sampling points, but compares the deviation of the current sampling point from the median of the previous and next windows, and judges whether the relevant variables change synchronously. When a single temperature point suddenly increases while the load factor, current and ambient temperature do not change accordingly, the temperature point is used as the sampling peak for boundary replacement. When multiple variables change synchronously according to electrothermal logic, the original change trajectory is retained. The multivariate operating state sequence obtained after the above processing can provide a data foundation for subsequent factor construction that is time-continuous, variable-corresponding and fluctuation-controlled. In this embodiment, the interference of asynchronous sampling, short-term missing and isolated peaks on the residual reconstruction results is reduced.

[0023] In this embodiment, the electrical excitation factor is constructed according to the relationship between changes in current, load rate, active power, and reactive power; the heat dissipation boundary factor is constructed according to the temperature difference relationship between ambient temperature, oil temperature, and winding temperature; and the thermal inertia response factor is constructed according to the hysteresis response of oil temperature and winding temperature relative to the electrical excitation. Specifically, the quantification method of the electrical excitation factor can be characterized by the following formula: ; in, This represents the electrical excitation factor at the t-th sampling time. Represents current. Indicates load rate, Indicates active power. Indicates reactive power. , and These represent the scale benchmarks for the corresponding changes within the historical normal operating range. This indicates a positive number used to avoid a denominator of zero. , and They represent , and The corresponding weights and their sums are 1; when Take 0.4, Take 0.3, When the value is 0.3 and the change in current increases synchronously with the change in load rate, The temperature increases with changes in electrical operating conditions, and the subsequent expected temperature value is allowed to rise within a reasonable lag range. When the temperature residual is small and continues to expand, the subsequent anomaly confidence index receives a higher input. In this embodiment, the instantaneous changes of multiple electrical quantities are compressed into the same excitation scale, which facilitates correlation calculation with the temperature hysteresis response.

[0024] Preferably, refer to Figure 2 Based on the multivariate operating state sequence, an electrical excitation factor, a heat dissipation boundary factor, and a thermal inertia response factor are constructed. This includes extracting the load change amplitude, load change slope, and power component change relationship from the current, load rate, active power, and reactive power to form the electrical excitation factor; extracting the heat dissipation temperature difference, environmental change trend, and temperature baseline offset from the ambient temperature, oil temperature, and winding temperature to form the heat dissipation boundary factor; and forming the thermal inertia response factor based on the hysteretic changes of the oil temperature and winding temperature relative to the electrical excitation factor. The electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor are then combined using the same time index to form an electrothermal coupling feature sequence. Specifically, the load change amplitude is determined by the current load rate and adjacent historical load rates. The absolute value of the load rate difference is determined, the load change slope is determined by the ratio of the load change amplitude to the sampling interval, the power component change relationship is determined by the direction and amplitude ratio between the active power change and the reactive power change, the heat dissipation temperature difference is jointly represented by the difference between oil temperature and ambient temperature and the difference between winding temperature and ambient temperature, the environmental change trend is represented by the sliding slope of the ambient temperature within a continuous window, the temperature baseline offset is represented by the offset of the current oil temperature and winding temperature relative to the median value of the historical normal temperature in the same load range, the hysteresis change is obtained by searching the candidate interval where the temperature change occurs earliest after the electrical excitation change, the same time index is used to ensure that the electrical excitation, heat dissipation boundary and thermal inertia response are jointly input at the same sampling time, and in this embodiment, the electrical quantity, temperature quantity and environmental quantity are avoided to be processed separately.

[0025] Further, forming the thermal inertia response factor includes determining candidate hysteresis intervals where the oil temperature and winding temperature change correspondingly, based on the change time of the electrical excitation factor; calculating the temperature rise slope, cumulative temperature rise, and temperature rise fall slope within the candidate hysteresis intervals; jointly normalizing the temperature rise slope, cumulative temperature rise, and temperature rise fall slope with the heat dissipation boundary factor to obtain a thermal response state vector; pairing the thermal response state vector within a continuous sampling period with the electrical excitation factor to form the thermal inertia response factor including load increase response, load hold response, and load decrease response; in specific implementation, the candidate hysteresis interval is determined by the direction change of the electrical excitation factor. As time progresses backward, the response of oil temperature and winding temperature within the candidate hysteresis interval is represented by monotonic segments, local peaks, and decline segments. The temperature rise slope reflects the rate of temperature increase after the start of the response, the cumulative temperature rise reflects the total temperature response within the hysteresis interval, and the temperature rise decline slope reflects the rate of temperature decline after the load decreases. During joint normalization, the heat dissipation boundary factor is used as the environmental reference for the temperature response, so that the reasonable temperature rise under high ambient temperature and the abnormal temperature rise under low ambient temperature have different calculation benchmarks. The paired thermal inertia response factor can describe the temperature response trajectory in the three types of operating segments: load increase, load hold, and load decrease. In this embodiment, whether the temperature change conforms to the hysteresis law of electrical excitation is transformed into a calculable state vector.

[0026] In a preferred embodiment, the thermal inertia response factor can be expressed in a windowed form using the following formula: ; in, This indicates that the excitation reference is the t-th sampling time and is delayed. Thermal inertial response after one sampling interval Indicates the winding temperature. Indicates oil temperature. Indicates ambient temperature. This represents the baseline for temperature difference within the historical normal operating range. , and They are respectively represented as , and The corresponding weights and their sums are 1; when When taking continuous sampling intervals of 1, 2, 3, etc., if the winding temperature and oil temperature gradually increase after the load increases, and the temperature difference remains within the boundary of the same operating condition, then Keeping the response close to historical normal levels, if the load change is small. If the value remains positive and the temperature difference term increases synchronously, then... If the response deviates from the normal response range, the subsequent residual coupling judgment will retain this abnormal trajectory. In this embodiment, the hysteresis length, temperature rise amplitude and heat dissipation boundary are incorporated into the same thermal response calculation formula.

[0027] Further, pairing the thermal response state vector with the electrical excitation factor within a continuous sampling period includes dividing the electrical excitation factor into segments according to the state transition order of load increase, load hold, and load decrease; extracting the initial value, peak value, fall value, and hysteresis length of the thermal response state vector within each segment; concatenating the initial value, peak value, fall value, and hysteresis length between adjacent segments to form a response memory matrix; using the response memory matrix as an additional input to the mechanism-constrained time-series reconstruction model, which, together with the electrothermal coupling feature sequence, participates in the generation of the desired state sequence; specifically, the load increase segment... The load holding segment is determined by the electrical excitation factor being continuously greater than the stable baseline within the same window. The load decreasing segment is determined by the electrical excitation factor being within a narrow fluctuation range. The load decreasing segment is determined by the electrical excitation factor being continuously less than the previous stable range. The initial value is used to record the baseline at the beginning of the temperature response. The peak value is used to record the local maximum point reached by the temperature hysteresis response. The fallback value is used to record the state of the temperature decreasing to the stable range after the load decreases. The hysteresis length is used to record the time interval between the change in electrical excitation and the change in temperature response. The response memory matrix enables the model to read the heat accumulation state left by the previous running segment when reconstructing the current expected temperature value. In this embodiment, it avoids misjudging the reasonable heat accumulation caused by the previous load as the current anomaly.

[0028] Preferably, refer to Figure 3The electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor are input into a mechanism-constrained time-series reconstruction model. This includes dividing the electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor into short-term impact windows, medium-term load fluctuation windows, and long-term temperature rise accumulation windows; encoding the variable change trajectories in different time windows to obtain multi-scale operating condition representations; setting constraint terms for the lag between electrical quantity changes and temperature quantity changes in the mechanism-constrained time-series reconstruction model; and inputting the multi-scale operating condition representations and the constraint terms together into the reconstruction calculation process to obtain the desired state sequence corresponding to the current operating condition. Specifically, the short-term impact window mainly retains the instantaneous change trajectories of voltage, current, and load rate; the medium-term load fluctuation window... The load fluctuation window mainly retains the continuous change trajectory of active power, reactive power, and load rate, while the long-term temperature rise accumulation window mainly retains the slow change trajectory between oil temperature, winding temperature, and ambient temperature. During encoding, the variables in each window are represented by sliding difference, mean offset, extreme value offset, and direction sequence. Then, the representation vectors of different windows are concatenated according to the same sampling time. The mechanism-constrained time-series reconstruction model can adopt a data reconstruction structure composed of an encoder and a decoder. The encoder receives multi-scale operating condition representation and response memory matrix, and the decoder outputs the expected values ​​of voltage, current, load rate, active power, reactive power, oil temperature, winding temperature, and ambient temperature. In this embodiment, short-term electrical shock and long-term temperature rise accumulation are processed separately in the same reconstruction process.

[0029] Furthermore, in the mechanism-constrained time-series reconstruction model, constraints on electrical quantity changes and temperature quantity lag changes are set, including generating lag response templates under different load ranges based on the correspondence between the electrical excitation factor and the thermal inertia response factor in the historical normal operation range; matching the real-time input multi-scale operating condition representation with the lag response templates to obtain response constraint weights; when reconstructing the expected values ​​of the oil temperature and the winding temperature, the response constraint weights are embedded in the reconstruction loss calculation, so that the expected state sequence simultaneously includes constraints on instantaneous electrical quantity changes and lag changes in temperature; in specific implementation, the historical normal... The normal operating range is divided into multiple sample clusters according to the load rate range and the ambient temperature range. Within each sample cluster, the hysteresis length, temperature rise slope, cumulative temperature rise, and fall slope of oil temperature and winding temperature after electrical excitation changes are statistically analyzed, thereby forming a hysteresis response template. The distance between the real-time input multi-scale operating condition characterization and each hysteresis response template is calculated. The smaller the distance, the higher the response constraint weight; the larger the distance, the lower the response constraint weight. The reconstruction loss not only calculates the numerical error between the actual value and the expected value, but also constrains the hysteresis direction and hysteresis amplitude of the expected temperature value relative to the electrical excitation. In this embodiment, the expected state sequence does not deviate from the normal electrothermal response law of the transformer.

[0030] In this embodiment, the training and real-time reconstruction of the mechanism-constrained temporal reconstruction model can be expressed using the following loss function: ; Where J represents the reconstruction loss, and r represents the number of variables involved in the reconstruction. This represents the actual value of the j-th variable at the t-th sampling time. This represents the expected value of the corresponding variable. Indicates the variable weights. and They represent lag respectively. The expected increments of winding temperature and oil temperature after each sampling interval This represents the change in the electrical excitation factor. and This represents the normal response coefficient of winding temperature and oil temperature relative to electrical excitation under operating condition category c. and Indicates the constraint strength; when the real-time operating condition is matched to a high-load sample cluster, and The expected temperature increment is allowed to rise within the hysteresis window. When the real-time operating condition matches the sample cluster of load decrease, the response coefficient limits the expected temperature value to continue to rise without constraint. In this embodiment, the numerical reconstruction and hysteresis response constraint are combined through a loss function.

[0031] Preferably, the step of generating an anomaly confidence index based on residual amplitude, residual duration, residual direction consistency, and intervariate residual coupling degree includes calculating the standardized deviation values ​​of electrical quantity residuals, temperature quantity residuals, and environmental quantity residuals respectively; determining the residual duration and residual direction consistency based on whether the residuals of the same variable maintain the same direction of deviation within a continuous sampling period; determining the intervariate residual coupling degree based on the matching relationship between electrical quantity residuals and temperature quantity residuals within a preset lag window; and weighting and fusing the standardized deviation values, residual duration, residual direction consistency, and intervariate residual coupling degree to obtain the anomaly confidence index. In specific implementation, the residual amplitude is determined by the actual operating state sequence and the expected... The residual sequence is obtained by subtracting each variable, and then standardized according to the residual scale of the historical normal operation interval. The duration of the residual is determined by the number of sampling periods that continuously exceed the normal residual boundary of the same operating condition. The consistency of the residual direction is determined by the proportion of the residual of the same variable that maintains a positive or negative deviation within a continuous window. The coupling degree of the intervariable residual is determined by the direction, slope and cumulative deviation relationship of the electrical quantity residual and the temperature quantity residual within the lag window. During weighted fusion, the temperature quantity residual has a higher weight under stable or decreasing load conditions, the electrical quantity residual has a higher weight under short-term impact conditions, and the environmental quantity residual is used to correct the heat dissipation boundary. In this embodiment, the early warning judgment considers the magnitude of the deviation, the duration of the deviation and the multivariable coupling relationship simultaneously.

[0032] Further, determining the intervariable residual coupling degree based on the matching relationship between electrical quantity residuals and temperature quantity residuals within a preset lag window includes: using the abrupt change point of the electrical quantity residual as the starting point of the window, and extracting temperature quantity residual subsequences at different lag step sizes; calculating the direction alignment ratio, slope correspondence ratio, and cumulative deviation ratio between the electrical quantity residual and each temperature quantity residual subsequence; determining the residual coupling vector based on the direction alignment ratio, the slope correspondence ratio, and the cumulative deviation ratio; and obtaining the intervariable residual coupling degree based on the distance between the residual coupling vector and the historical normal coupling vector. Specifically, the abrupt change point of the electrical quantity residual is determined by the current residual, load rate residual, and... The joint changes in active power residuals and reactive power residuals are determined. The temperature residual subsequence includes the winding temperature residual subsequence and the oil temperature residual subsequence. The direction consistency ratio indicates the proportion of sampling points whose direction of change of electrical quantity residuals is the same as the direction of change of temperature quantity residuals. The slope correspondence ratio indicates the normalized closeness between the slope of change of electrical quantity residuals and the slope of change of temperature quantity residuals. The cumulative deviation ratio indicates the correspondence between the cumulative amount of temperature residuals and the cumulative amount of electrical quantity residuals within the lag window. The historical normal coupling vector is statistically obtained from the same load and environmental conditions in the normal operating range. In this embodiment, whether the temperature change follows the electrical excitation is transformed into a vector distance judgment in the residual space.

[0033] In this embodiment, the anomaly confidence index can be calculated using the following formula: ; in, Let represent the anomaly confidence index at the t-th sampling time. This represents the normalized result of the residual amplitude. This represents the normalized result of the residual duration. This indicates the consistency of the residual direction. Indicates the degree of intervariate residual coupling. , , and They are respectively represented as , , and The corresponding weights and their sums are 1; when Take 0.3, Take 0.2, Take 0.2, When the value is 0.3, the short-term spike is due to Smaller and When there is no sustained deviation and it is difficult to form a high confidence index independently, and the temperature residual remains positive after the load decreases and the intervariate residual coupling relationship deviates from the normal vector, , and Together increase and promote Entering a higher warning level; the weighting terms in the formula can be set according to the residual stability of the historical normal operation interval, but each weight maintains the same meaning within the same deployment cycle. In this embodiment, different residual characteristics are unified to a numerical scale that can be graded and judged.

[0034] In a preferred embodiment, the intervariate residual coupling degree can be calculated using the following formula: ; in, This represents the set of allowed lag steps for searching. This represents the residual coupling vector composed of the proportion of direction alignment, the proportion of slope correspondence, and the proportion of cumulative deviation of electrical quantity residuals. Indicates lag The residual coupling vector, formed by the proportion of the direction of the temperature residual, the proportion of the slope, and the proportion of the cumulative deviation after each sampling interval, is a sample of temperature residuals. Represents the 2-norm; when If we take several consecutive lag steps, and the distance between the two types of vectors is small at a certain lag step, it indicates that the temperature residual can respond to the electrical residual according to the normal lag relationship. A smaller value indicates that the distance between the two types of vectors is large across all hysteresis steps, meaning the temperature residual fails to match the electrical residual within the allowable hysteresis range. With larger values ​​and participation in the calculation of the outlier confidence index, this embodiment does not require a fixed single lag time to evaluate the intervariate residual relationship.

[0035] Furthermore, the adaptive early warning boundary based on the historical normal operating range includes dividing the historical normal operating range into boundary sample clusters according to load range and ambient temperature range; calculating the quantile boundary of the anomaly confidence index and the stability boundary of the intervariate residual coupling degree in each boundary sample cluster; matching the load range and ambient temperature range corresponding to the real-time data to the target boundary sample cluster, and interpolating according to the boundary differences of adjacent boundary sample clusters to obtain the adaptive early warning boundary under the current operating condition; comparing the anomaly confidence index with the adaptive early warning boundary to generate the transformer real-time data anomaly early warning result; in specific implementation, the load range can be based on the load rate in the historical normal operating data. The distribution is divided according to the ambient temperature range. Each boundary sample cluster stores the normal upper boundary of the abnormal confidence index under the same operating condition, the normal fluctuation boundary of the intervariable residual coupling degree, and the normal persistence boundary of the temperature residual. After the real-time data arrives, the sample cluster to which it belongs is determined by the current load rate and ambient temperature. When the real-time operating condition is between adjacent sample clusters, boundary interpolation is performed according to the distance instead of directly using a fixed threshold. When classifying the judgment, the state where the abnormal confidence index is lower than the current operating condition boundary is retained as normal monitoring, the state that is continuously close to the boundary is marked as attention, and the state that exceeds the boundary and persists is marked as abnormal warning. In this embodiment, the warning boundary is kept in correspondence with the current load and environmental conditions.

[0036] In this embodiment, the adaptive early warning boundary can be expressed by the following formula: ; in, This indicates the adaptive warning boundary under the current operating conditions. , , and These represent the boundary sample clusters corresponding to adjacent load ranges and adjacent ambient temperature ranges, respectively. This represents the normalized position of the real-time load rate between adjacent load intervals, and v represents the normalized position of the real-time ambient temperature between adjacent ambient temperature intervals. When the real-time load rate is close to the m-th load interval and the ambient temperature is close to the (n+1)-th ambient temperature interval... Smaller and v is larger right The contribution is significant, and the boundary calculation results are closer to the normal temperature residual distribution under the current environment. When the load rate and ambient temperature are both in the middle of adjacent intervals, the four boundary sample clusters participate in the calculation together to avoid the boundary from jumping with the interval switching. In this embodiment, the graded early warning boundary changes continuously and retains the basis for historical normal operation.

[0037] In one embodiment, reference Figure 4This invention provides a real-time data anomaly early warning system for transformers, comprising: a data processing module for acquiring real-time data on voltage, current, load rate, active power, reactive power, oil temperature, winding temperature, and ambient temperature during transformer operation, and performing alignment, missing value compensation, and spike suppression according to sampling timestamps to obtain a multivariate operating state sequence; an electrothermal coupling feature construction module for constructing electrical excitation factors, heat dissipation boundary factors, and thermal inertia response factors based on the multivariate operating state sequence; a mechanism constraint residual reconstruction module for generating a desired state sequence under normal electrothermal coupling conditions based on the electrical excitation factors, heat dissipation boundary factors, and thermal inertia response factors, and calculating the residual sequence between the multivariate operating state sequence and the desired state sequence; and an anomaly confidence calculation module for calculating the residual confidence based on residual amplitude, residual duration, residual direction consistency, and cross-variable... The residual coupling degree generates an anomaly confidence index; the early warning judgment module is used to classify and judge the anomaly confidence index based on the adaptive early warning boundary formed by the historical normal operation range and output the real-time data anomaly early warning result of the transformer; in specific implementation, the data processing module receives real-time data from the transformer operation acquisition link and converts data from different sampling sources into a multivariate operating state sequence under a unified time index; the electrothermal coupling feature construction module extracts electrical excitation, heat dissipation boundary and thermal inertia response according to the same time index; the mechanism constraint residual reconstruction module generates the expected state sequence based on the multi-scale operating condition characterization and hysteresis response template; the anomaly confidence calculation module calculates the amplitude, duration, direction and coupling characteristics in the residual space; the early warning judgment module matches the adaptive early warning boundary according to the current operating condition and forms the output. In this embodiment, the system-side processing flow and the method-side calculation logic are consistent.

[0038] In this embodiment, the data processing module includes a data access process, a time index cache, a missing data compensation process, and a peak suppression process. The data access process receives real-time data according to three types of fields: variable name, sampling time, and sampled value. The time index cache establishes a row index according to the sampling time and writes voltage, current, load rate, active power, reactive power, oil temperature, winding temperature, and ambient temperature into the corresponding columns. The missing data compensation process uses adjacent time constraints to compensate for short-term missing variables and uses historical baselines and related variable trends to compensate for continuous missing variables. The peak suppression process performs median window testing on univariate abrupt changes and concludes the analysis. The synchronization of relevant variables is used to determine whether to replace them; the electrothermal coupling feature construction module includes an electrical excitation calculation process, a heat dissipation boundary calculation process, and a thermal inertial response calculation process. The electrical excitation calculation process reads the change amplitude and slope of current, load rate, active power, and reactive power. The heat dissipation boundary calculation process reads the temperature difference and baseline offset between oil temperature, winding temperature, and ambient temperature. The thermal inertial response calculation process searches for candidate hysteresis intervals of temperature response after the electrical excitation changes and generates a thermal response state vector. Each process uses the same time index to transmit data. In this embodiment, time misalignment between front-end data processing and feature construction is avoided.

[0039] In this embodiment, the mechanism-constrained residual reconstruction module includes a window partitioning process, a multi-scale encoding process, a template matching process, a desired state generation process, and a residual output process. The window partitioning process divides the electrothermal coupling feature sequence into a short-term impact window, a medium-term load fluctuation window, and a long-term temperature rise accumulation window. The multi-scale encoding process extracts the variable change trajectory within each window and forms a multi-scale operating condition representation. The template matching process selects a hysteresis response template from the historical normal operating range based on the real-time load range and the ambient temperature range and calculates the response constraint weights. The desired state generation process inputs the multi-scale operating condition representation, the response memory matrix, and the response constraint weights into the reconstruction calculation process and outputs the desired state sequence under normal electrothermal coupling conditions. The residual output process subtracts the actual multivariate operating state sequence from the desired state sequence variable by variable and outputs the residual sequence. The mechanism-constrained residual reconstruction module completes one reconstruction in each real-time window and passes the response memory matrix of the previous window to the next window, so that the heat accumulation in continuous operating condition transitions is preserved. In this embodiment, it is ensured that the residual sequence comes from the desired state under the same mechanism constraint rather than an isolated prediction result.

[0040] In a preferred embodiment, the anomaly confidence calculation module includes a residual standardization process, a duration statistics process, a direction consistency statistics process, and a cross-variable residual coupling process. The residual standardization process converts electrical quantity residuals, temperature quantity residuals, and environmental quantity residuals into comparable values ​​according to the residual scale in the historical normal operation interval. The duration statistics process records the number of sampling periods in which the residual of the same variable continuously exceeds the normal boundary of the same operating condition. The direction consistency statistics process determines whether the residual direction in a continuous window remains in the same direction. The cross-variable residual coupling process uses the abrupt change point of the electrical quantity residual as the starting point of the window and truncates the temperature quantity residual at different lag step sizes. The residual subsequence is measured, and then the proportion of direction consistency, slope correspondence, and cumulative deviation are calculated. The early warning determination module includes a boundary sample cluster matching process, a boundary interpolation process, and a hierarchical output process. The boundary sample cluster matching process determines the target boundary sample cluster based on the real-time load rate and ambient temperature. The boundary interpolation process calculates the adaptive early warning boundary under the current operating conditions based on the differences between adjacent sample clusters. The hierarchical output process outputs the normal monitoring, attention status, or abnormal early warning status based on the comparison result between the abnormal confidence index and the adaptive early warning boundary. In this embodiment, residual calculation, boundary matching, and hierarchical determination are separated into continuous data processing links.

[0041] In this embodiment, the establishment of historical normal operation intervals does not depend on fault samples. Specifically, time periods that have been manually reviewed or confirmed by operation records as having no abnormal alarms can be selected as initial normal samples. After forming sample clusters according to load intervals and ambient temperature intervals, the distribution of electrical excitation factors, heat dissipation boundary factors, thermal inertia response factors, residual amplitude distribution, intervariate residual coupling vector distribution, and anomaly confidence index boundaries are saved for each sample cluster. After new real-time data is determined to be normal monitoring and remains stable, it can be written to the candidate normal cache according to the operation date. The candidate normal cache is only included in the boundary sample cluster update when the residual amplitude, residual direction consistency, and intervariate residual coupling degree are all within the boundary of the target sample cluster. When updating the boundary sample cluster, the statistical benchmark of the original normal samples is retained and the new samples are given restricted weights to avoid short-term unidentified anomalies being directly absorbed as normal boundaries. In this embodiment, the adaptive warning boundary has the ability to update operation data without changing the electrothermal coupling judgment logic.

[0042] In a preferred embodiment, the early warning result output includes anomaly level, trigger sampling time, anomaly confidence index, corresponding adaptive early warning boundary, main residual variables, and intervariate residual coupling degree. The output content is extracted by the early warning judgment module from the calculation results of the same real-time window, without adding business judgments unrelated to transformer operation. When the anomaly confidence index exceeds the adaptive early warning boundary but the residual duration is short, the attention status is output and the subsequent window tracking mark is retained. When the anomaly confidence index exceeds the adaptive early warning boundary and the residual direction consistency and intervariate residual coupling degree are both in the abnormal range, the anomaly early warning status is output. When the anomaly confidence index falls back to below the current operating condition boundary and the continuous window residuals no longer deviate in the same direction, the tracking mark is canceled and the residual trajectory of that segment is saved for boundary sample cluster verification. The output process only uses the calculation results formed by voltage, current, load rate, active power, reactive power, oil temperature, winding temperature, and ambient temperature, and does not introduce external variables unrelated to real-time data anomaly early warning. In this embodiment, it is ensured that the early warning conclusion can be traced back to the multivariate operating state sequence, expected state sequence, and residual sequence.

Claims

1. A method for real-time data anomaly early warning of transformers, characterized in that, include: Real-time data of voltage, current, load rate, active power, reactive power, oil temperature, winding temperature and ambient temperature during transformer operation are acquired. The real-time data is aligned, missing data is compensated and spikes are suppressed according to the sampling timestamp to obtain a multivariable operating state sequence. Based on the multivariate operating state sequence, an electrical excitation factor, a heat dissipation boundary factor, and a thermal inertia response factor are constructed. The electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor are then input into a mechanism-constrained time-series reconstruction model to generate the desired state sequence under normal electrothermal coupling conditions. Calculate the residual sequence between the multivariate running state sequence and the desired state sequence, and generate an anomaly confidence index based on residual magnitude, residual duration, residual direction consistency and intervariate residual coupling degree; Based on the adaptive early warning boundary formed by the historical normal operation range, the abnormal confidence index is classified and judged, and the real-time data abnormality early warning result of the transformer is output.

2. The method for real-time data anomaly early warning of transformers according to claim 1, characterized in that, Constructing electrical excitation factors, heat dissipation boundary factors, and thermal inertia response factors based on the multivariate operating state sequence includes: extracting the load change amplitude, load change slope, and power component change relationship from the current, the load rate, the active power, and the reactive power to form the electrical excitation factors; The heat dissipation temperature difference, environmental change trend, and temperature baseline shift are extracted from the ambient temperature, the oil temperature, and the winding temperature to form the heat dissipation boundary factor. The thermal inertia response factor is formed based on the hysteretic changes in the oil temperature and the winding temperature relative to the electrical excitation factor. The electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor are combined into an electrothermal coupling feature sequence according to the same time index.

3. The method for real-time data anomaly early warning of transformers according to claim 1, characterized in that, The electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor are input into the mechanism-constrained time-series reconstruction model, including: dividing the electrical excitation factor, the heat dissipation boundary factor, and the thermal inertia response factor into a short-term impact window, a medium-term load fluctuation window, and a long-term temperature rise accumulation window; The variable change trajectories in different time windows are encoded to obtain multi-scale working condition representations; In the mechanism-constrained time-series reconstruction model, constraint terms are set for the lag changes in electrical quantities and temperature quantities. The multi-scale operating condition characterization and the constraint terms are input into the reconstruction calculation process to obtain the desired state sequence corresponding to the current operating condition.

4. The method for real-time data anomaly early warning of transformers according to claim 1, characterized in that, The generation of anomaly confidence indices based on residual amplitude, residual duration, residual direction consistency, and intervariate residual coupling degree includes: calculating the standardized deviation values ​​of electrical quantity residuals, temperature quantity residuals, and environmental quantity residuals, respectively. The duration of the residual and the consistency of the residual direction are determined based on whether the residuals of the same variable maintain the same direction of deviation within consecutive sampling periods. The degree of coupling of the intervariable residuals is determined based on the matching relationship between the electrical quantity residuals and the temperature quantity residuals within a preset hysteresis window. The anomaly confidence index is obtained by weighting and fusing the standardized deviation, the residual duration, the residual directional consistency, and the intervariate residual coupling degree.

5. The method for real-time data anomaly early warning of a transformer according to claim 2, characterized in that, The formation of the thermal inertia response factor includes: using the change time of the electrical excitation factor as a reference, determining the candidate hysteresis intervals where the oil temperature and the winding temperature show corresponding changes; Calculate the temperature rise slope, cumulative temperature rise, and temperature rise fall slope within the candidate hysteresis interval; The temperature rise slope, the cumulative temperature rise, and the temperature rise fall slope are jointly normalized with the heat dissipation boundary factor to obtain the thermal response state vector. The thermal response state vector within a continuous sampling period is paired with the electrical excitation factor to form the thermal inertial response factor, which includes the load increase response, the load hold response, and the load decrease response.

6. The method for real-time data anomaly early warning of a transformer according to claim 3, characterized in that, In the mechanism-constrained time-series reconstruction model, constraints on the hysteretic changes of electrical quantities and temperature quantities are set, including: generating hysteretic response templates under different load ranges based on the correspondence between the electrical excitation factor and the thermal inertia response factor in the historical normal operation range; The multi-scale working condition representation input in real time is matched with the hysteresis response template to obtain the response constraint weights; When reconstructing the expected values ​​of the oil temperature and the winding temperature, the response constraint weights are embedded in the reconstruction loss calculation, so that the expected state sequence simultaneously includes constraints on instantaneous changes in electrical quantities and constraints on delayed changes in temperature quantities.

7. The method for real-time data anomaly early warning of transformers according to claim 4, characterized in that, The step of determining the intervariable residual coupling degree based on the matching relationship between the electrical quantity residual and the temperature quantity residual within a preset lag window includes: taking the abrupt change point of the electrical quantity residual as the starting point of the window, and extracting the temperature quantity residual subsequence under different lag step sizes; Calculate the proportion of direction alignment, slope correspondence, and cumulative deviation between the electrical quantity residual and each temperature quantity residual subsequence; The residual coupling vector is determined based on the direction alignment ratio, the slope correspondence ratio, and the cumulative deviation ratio. The intervariable residual coupling degree is obtained based on the distance between the residual coupling vector and the historical normal coupling vector.

8. The method for real-time data anomaly early warning of a transformer according to claim 5, characterized in that, The step of pairing the thermal response state vector with the electrical excitation factor within a continuous sampling period includes: dividing the electrical excitation factor into segments according to the state transition order of load increase, load hold and load decrease; Extract the initial value, peak value, fall value, and hysteresis length of the thermal response state vector within each segment; The initial value, peak value, fallback value and hysteresis length between adjacent segments are sequentially concatenated to form a response memory matrix; The response memory matrix is ​​used as an additional input to the mechanism-constrained temporal reconstruction model, and together with the electrothermal coupling feature sequence, it participates in the generation of the desired state sequence.

9. A method for real-time data anomaly early warning of a transformer according to claim 7, characterized in that, The adaptive early warning boundary based on the historical normal operation range includes: dividing the boundary sample clusters within the historical normal operation range according to the load range and the ambient temperature range; Calculate the quantile boundaries of the anomaly confidence index and the stability boundaries of the intervariate residual coupling degree in each boundary sample cluster; The load range and ambient temperature range corresponding to the real-time data are matched to the target boundary sample cluster, and interpolation is performed according to the boundary difference of adjacent boundary sample clusters to obtain the adaptive warning boundary under the current working condition. The anomaly confidence index is compared with the adaptive early warning boundary to generate the real-time data anomaly early warning result for the transformer.

10. A real-time data anomaly early warning system for transformers, characterized in that, include: The data processing module is used to acquire real-time data on voltage, current, load rate, active power, reactive power, oil temperature, winding temperature and ambient temperature during transformer operation, and to perform alignment, missing data compensation and spike suppression according to the sampling timestamp to obtain a multivariable operating state sequence. An electrothermal coupling feature construction module is used to construct electrical excitation factors, heat dissipation boundary factors, and thermal inertia response factors based on the multivariate operating state sequence. The mechanism-constrained residual reconstruction module is used to generate the desired state sequence under normal electrothermal coupling state based on the electrical excitation factor, the heat dissipation boundary factor and the thermal inertia response factor, and to calculate the residual sequence between the multivariate operating state sequence and the desired state sequence. The anomaly confidence calculation module is used to generate anomaly confidence indices based on residual magnitude, residual duration, residual direction consistency, and intervariate residual coupling degree. The early warning determination module is used to classify and determine the abnormal confidence index based on the adaptive early warning boundary formed by the historical normal operation range and output the real-time data abnormality early warning result of the transformer.

Images