A method and device for intelligent monitoring of transformer operating energy efficiency

Abstract

This invention discloses a method and device for intelligent monitoring of transformer operating energy efficiency, comprising: acquiring multi-dimensional time-series data of transformer iron loss, copper loss, etc.; generating a cross-domain influence factor or cross-domain influence matrix to quantify the cross-domain correlation strength between the iron loss and copper loss time series by calculating the cross-correlation function between the two; and integrating the quantitative correlation feature with standardized loss trajectories and operating parameters to construct a machine learning prediction model to identify and warn of the hidden fault mode of "iron loss-copper loss cross-domain correlation degradation" driven by iron loss deterioration. This invention overcomes the shortcomings of traditional technology that analyzes losses in isolation, and can achieve early warning and source tracing of specific composite faults, significantly improving the safety and reliability of transformer operation. This invention also provides a device for executing the method.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring and diagnosis technology, and in particular to a method and device for intelligent monitoring of transformer operating energy efficiency. Background Technology

[0002] Transformers are key equipment in power systems for transmitting and transforming electrical energy. Their operational efficiency and reliability are crucial for ensuring the safe and stable operation of the power grid and improving energy utilization efficiency. The performance and efficiency of a transformer are primarily affected by two core losses: iron losses and copper losses. Iron losses, also known as no-load losses, are mainly caused by the hysteresis and eddy current effects within the transformer core and remain essentially constant even when the transformer is unloaded. Copper losses, also known as load losses, are the heat generated by the resistance of the copper windings when current flows through them; their magnitude is proportional to the square of the load current. These two types of losses not only directly convert into heat energy, reducing the transformer's energy conversion efficiency, but the heat generated also leads to an increase in the internal temperature of the transformer, accelerating the aging of the insulation materials and thus affecting its operational reliability and service life. In actual operation, iron losses and copper losses coexist and are not entirely independent. For example, the increased core temperature due to increased iron losses affects the winding temperature through heat conduction, thereby changing the winding resistance and influencing copper losses, forming a complex interaction.

[0003] Currently, transformer monitoring technologies mostly focus on monitoring single physical quantities such as temperature, dissolved gases in oil, and partial discharge, or on isolated analyses of performance parameters such as load losses (copper losses) and no-load losses (iron losses). These methods exhibit certain limitations when dealing with transformer performance degradation under specific complex operating conditions. For example, existing technologies often treat iron losses and copper losses as independent variables, ignoring the potential physical correlation between them. This makes it difficult to identify composite fault modes, namely, cross-domain correlated degradation of iron losses and copper losses (hereinafter referred to as "correlated degradation").

[0004] This type of degradation is particularly prevalent in thermally coupled compact transformers, such as dry-type epoxy resin cast transformers or oil-immersed transformers whose heat dissipation capacity is reduced due to poor internal oil circulation. In these transformers, the core and windings form a tight thermodynamic coupling due to their compact structure or obstructed heat dissipation paths. When the transformer experiences an abnormal increase in iron losses due to factors such as material aging, harmonic pollution, or overvoltage, the extra heat generated by the core cannot be effectively dissipated through conventional paths. Instead, it forms a localized heat concentration and conduction path (hereinafter referred to as the "thermal bridge" effect) between the core and windings through solid-state heat conduction or restricted convection, exerting continuous thermal stress on adjacent windings. This cross-physical-domain heat transfer triggers a series of chain reactions: firstly, it causes the resistivity of the windings to exceed the expected value at normal operating temperatures, resulting in an abnormal increase in copper losses even when the load current has not changed significantly; secondly, continuous thermal stress accelerates the deterioration of the physical and chemical properties of the winding insulation materials, such as embrittlement of the insulating varnish and the formation of microcracks, laying the groundwork for serious electrical faults such as inter-turn short circuits.

[0005] Existing technical solutions are generally unable to effectively identify the abrupt change in copper loss from "abnormal increase in iron loss" to "abnormal change in copper loss". In the early stages of fault development, changes in copper loss may not be obvious, making it undetectable by the monitoring system; however, when a significant increase in copper loss is detected, it usually indicates that irreversible damage has occurred to the winding insulation, or even that it is on the verge of breakdown. At this point, intervention is unlikely to prevent significant losses. Summary of the Invention

[0006] To address the aforementioned technical problems, one aspect of the present invention provides an intelligent monitoring method for transformer operating efficiency. This method quantifies the dynamic correlation strength between iron loss and copper loss and incorporates it as a key feature into a prediction model, thereby accurately identifying potential degradation risks driven by iron loss deterioration. The method includes the following steps:

[0007] Step S1 involves acquiring and constructing multi-dimensional time-series data. This step obtains operational data from the transformer online monitoring system or historical database using a preset fixed sampling frequency. Specifically, this step simultaneously acquires iron loss and copper loss measurement data reflecting the transformer's loss status, as well as load current monitoring values, operating temperature data of key components, and voltage level records from the grid side reflecting its operating conditions. Each dimension of data is accompanied by a precise timestamp to construct a multi-dimensional time-series dataset. This dataset is stored according to parameter properties, such as loss parameter sequences and operating condition parameter sequences, providing a structured, high-fidelity data foundation for subsequent dynamic analysis and correlation modeling.

[0008] Step S2 involves dynamic identification of the operating mode. This step identifies the main sources of loss based on the transformer's actual load, providing a basis for subsequent targeted analysis. Specifically, based on the load current monitoring values ​​in the multi-dimensional time-series dataset constructed in Step S1, the operating mode of the transformer at each time point is determined to identify either the copper loss-dominated mode or the iron loss-dominated mode. Through this step, a corresponding operating mode identifier is generated for each time point, and the corresponding operating parameters are extracted to form an operating parameter input sequence.

[0009] Step S3 involves independent loss trajectory anomaly analysis. Based on the operating mode identified in Step S2, this step standardizes and detects anomalies in the dominant loss historical data under this mode to initially screen out time periods that may show signs of degradation. Specifically, under the copper loss-dominated mode, the copper loss value measurement data sequence is processed using the z-score normalization method, calculated as z = (x - μ) / σ. Here, x is the current data value, μ is the mean of the data sequence, and σ is the standard deviation of the data sequence. Subsequently, the rate of change of the normalized copper loss sequence is calculated, and time periods with rates exceeding a preset copper loss rate of change threshold are extracted as independent copper loss degradation samples. Under the iron loss-dominated mode, the iron loss value measurement data sequence is also processed using the z-score normalization method to calculate its rate of change. Time periods where iron loss shows a continuous upward trend and the rate of increase exceeds a preset iron loss growth rate threshold are specifically marked as "potential associated degradation source" time periods.

[0010] Step S4 involves a quantitative assessment of the cross-domain correlation between iron loss and copper loss. This step is crucial for achieving the objectives of this invention. It initiates in-depth cross-domain correlation analysis on the "potential correlation degradation source" time period marked in step S3 to quantitatively assess the driving force of iron loss anomalies on copper loss.

[0011] After identifying the "potentially correlated degradation source" period, the system extracts the time series of iron loss and copper loss change rates within that period. Subsequently, the cross-correlation function between these two time series is calculated, which can be expressed as: Cross-correlation function R_xy(τ) = E[(X_t - μ_x)(Y_{t+τ} - μ_y)]. In this formula, X_t and Y_t are the iron loss and copper loss change rate series, respectively; μ_x and μ_y are the means of series X_t and Y_t, respectively; τ is the time delay; and E represents the expected value. The calculated result is a function of the time delay τ. The magnitude of the peak value of this function represents the maximum correlation strength between the two series at a certain time delay, and the x-axis corresponding to the peak value is the maximum delay time τ_peak. This invention defines the peak value of this cross-correlation function as the "Cross-domain Influence Factor (CDIF)".

[0012] Step S5 involves performing degradation prediction modeling based on fused correlation features. This step constructs a prediction model capable of identifying multiple failure modes, including independent degradation and correlated degradation. Specifically, a training sample set fusing multidimensional features is constructed, where each sample includes: the standardized loss trajectory for the corresponding time period, the corresponding operating condition parameter input sequence, and the cross-domain influence factor calculated in step S4. A machine learning algorithm is used to construct the fused degradation prediction model, which can learn and identify the "iron loss-copper loss cross-domain correlated degradation" mode characterized by high CDIF values.

[0013] Step S6 involves tiered early warning and risk tracing. This step interprets and applies the model's prediction results, providing early warning information with practical guidance. First, the degradation tendency probability values ​​output by the model in Step S5 are classified into risk levels. When the judgment result is high or medium risk, an early warning is triggered. Simultaneously, to achieve accurate risk tracing, the current operating parameters and calculated cross-domain correlation features are obtained, and the similarity between their feature vectors and the feature vectors of high degradation risk samples already labeled in the historical database is calculated to generate refined identification results that include degradation pattern types.

[0014] In another aspect, the present invention provides a transformer operating energy efficiency intelligent monitoring device for performing the above-described method. The device may include:

[0015] A data acquisition module is used to acquire iron loss measurement data and copper loss measurement data from the transformer operation monitoring system at a fixed sampling frequency, simultaneously acquire load current monitoring values, operating temperature acquisition values ​​and voltage level record values, and construct a multi-dimensional time-series structured dataset containing timestamps;

[0016] An operation mode recognition module is used to determine the transformer operation mode based on the load current monitoring value in the multi-dimensional time-series structure dataset, and obtain the corresponding operation mode identifier and operating condition parameter input sequence;

[0017] An independent loss analysis module is used to standardize historical data of dominant loss according to the operating mode, extract independent degradation samples, and mark the time period of "potential associated degradation source".

[0018] A cross-domain correlation assessment module is used to calculate the cross-correlation function between iron loss-related time series and copper loss-related time series for the "potential correlation degradation source" period, so as to generate a cross-domain influence factor (CDIF) for quantifying the cross-domain correlation strength.

[0019] A predictive modeling module is used to construct a degradation prediction model that integrates standardized loss trajectories, operating parameters, and the cross-domain influencing factors, and outputs a degradation tendency probability value.

[0020] An early warning and source tracing module is used to classify the risk level of the degradation tendency probability value and perform similarity analysis with historical samples to trigger an early warning and generate refined identification results containing degradation pattern types.

[0021] The device can be implemented through hardware circuitry or through software programs running on general-purpose computing devices (such as servers and industrial control computers). Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0023] Figure 1 A flowchart of a method for intelligent monitoring of transformer operating energy efficiency provided in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the cross-correlation function derivation process in an embodiment of the present invention, wherein (a) represents two time series with delay correlation, and (b) represents the cross-correlation function and peak value (CDIF).

[0025] Figure 3 This is a structural block diagram of a transformer operation energy efficiency intelligent monitoring device provided in an embodiment of the present invention.

[0026] Figure 4 This is a comparison diagram of the changes in sample separability in the feature space before and after the introduction of the cross-domain influence factor in an embodiment of the present invention. (a) is the traditional feature space, and (b) is the feature space fused with the CDIF of the present invention.

[0027] Figure 5This is a comparison diagram of the cross-domain influence matrix heatmap under different health states in the embodiments of the present invention, where (a) is the CDIM under the healthy state and (b) is the CDIM under the associated degradation state.

[0028] Figure 6 This is a schematic diagram of one embodiment of the process of associated degradation fault evolution, monitoring and graded early warning in the present invention, wherein (a) is the original loss data trajectory, (b) is the evolution of cross-domain impact factor (CDIF), and (c) is the output risk level and early warning.

[0029] Figure 7 This is a schematic diagram illustrating the principle of risk tracing and degradation pattern identification based on cosine similarity in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0031] Example 1

[0032] This invention provides an intelligent monitoring method for transformer operating efficiency. By deeply analyzing the dynamic correlation between iron loss and copper loss, it identifies and warns of the "cross-domain correlation degradation of iron loss and copper loss" fault mode. (Refer to...) Figure 1 The method includes the following steps:

[0033] Step S1 involves acquiring and constructing multi-dimensional time-series data. This step obtains operational data from the transformer online monitoring system or historical database using a preset fixed sampling frequency. Specifically, this step simultaneously acquires iron loss and copper loss measurement data reflecting the transformer's loss status, as well as load current monitoring values, operating temperature data of key components, and voltage level records from the grid side reflecting its operating conditions. Each dimension of data is accompanied by a precise timestamp to construct a multi-dimensional time-series dataset. This dataset is stored according to parameter properties, such as loss parameter sequences and operating condition parameter sequences, providing a structured, high-fidelity data foundation for subsequent dynamic analysis and correlation modeling.

[0034] Specifically, this step can be broken down into three sub-steps: data acquisition, data synchronization, and structured storage. In the data acquisition sub-step, the load current can be measured using current transformers (CTs) or Rogowski coils installed on the high-voltage or low-voltage side of the transformer; the operating temperature can be collected using thermocouples or platinum resistance thermometers (RTDs) placed in the core, winding surfaces, and cooling medium (such as transformer oil); voltage level records are typically obtained from the Supervisory Control and Data Acquisition (SCADA) system; and iron loss and copper loss values ​​can be measured online using a high-precision power quality analyzer, or indirectly calculated based on basic electrical data such as voltage, current, and power factor, according to relevant electrical engineering standards (such as IEEE Std C57.12.90). Preferably, the sampling frequency can be set to once every 5 minutes to ensure data granularity while avoiding excessive burden on data storage and processing. For example, for a 10kV / 0.4kV dry-type epoxy resin cast transformer deployed in a data center, its data acquisition system continuously records its three-phase current, winding temperature, core temperature, input voltage, iron loss power, and copper loss power at 5-minute intervals. This data, along with timestamps, is stored in a time-series database (such as InfluxDB), forming records in the format {timestamp, I_A, I_B, I_C, T_winding (winding temperature), T_core, V_in (input voltage), P_fe (iron loss power), P_cu (copper loss power)}. This step provides the foundation for subsequent analysis, and its data quality directly determines the accuracy of fault identification. As an alternative, a non-fixed-frequency event-driven sampling strategy can be adopted. For example, when the load current change rate exceeds a preset threshold, the sampling frequency can be temporarily increased to capture the dynamic process more precisely.

[0035] Step S2 involves dynamic identification of the operating mode. This step identifies the main sources of loss based on the actual load of the transformer, providing a basis for subsequent targeted analysis. Specifically, based on the load current monitoring values ​​in the multi-dimensional time-series dataset constructed in Step S1, the operating mode of the transformer at each time point is determined. If the load current monitoring value continuously exceeds 80% of the transformer's rated current within a certain time period, that period is identified as the copper loss-dominated mode; if the load current monitoring value continuously falls below 30% of the rated current, that period is identified as the iron loss-dominated mode. Through this step, a corresponding operating mode identifier is generated for each time point, and the corresponding operating parameters (such as load current, temperature, and voltage) are extracted to form an operating parameter input sequence.

[0036] This step can be implemented using a rule-based classifier. For example, input a load current sequence within a time window (e.g., 24 hours) and calculate the ratio of its average value (I_avg) to the transformer's rated current I_N. If I_avg / I_N > 0.8, the window is marked as "copper loss dominant"; if I_avg / I_N < 0.3, it is marked as "iron loss dominant". Periods between 0.3 and 0.8 can be marked as "mixed mode" and can be temporarily ignored or subjected to different analysis strategies in subsequent analyses. The principle behind this step is that the total transformer loss consists of a basically constant iron loss and copper loss that varies with the square of the load current. Under high load, copper loss is much greater than iron loss, becoming the key to analyzing the winding health; under low load, copper loss is very small, and iron loss becomes the main component of the total loss, making the analysis of iron loss changes most clear. This mode division can effectively decouple the loss characteristics under different operating conditions, avoiding strong interference from one loss while analyzing another, thereby improving the signal-to-noise ratio of the analysis. As an alternative, unsupervised learning algorithms, such as K-means clustering, can be used to cluster operating data characterized by load current, power factor, etc., and automatically discover the typical operating modes of transformers. This approach is more adaptive than the fixed threshold method.

[0037] Step S3 involves performing independent loss trajectory anomaly analysis. Based on the operating mode identified in Step S2, this step standardizes and detects anomalies in the dominant loss historical data under this mode to initially screen out time periods that may show signs of degradation. Specifically, under the copper loss-dominant mode, the copper loss value measurement data sequence is processed using the z-score normalization method, calculated as z = (x - μ) / σ. Here, x is the current data value, μ is the mean of the data sequence, and σ is the standard deviation of the data sequence. Subsequently, the rate of change of the normalized copper loss sequence is calculated, and time periods where the rate of change exceeds a preset copper loss rate of change threshold (e.g., the absolute value of the rate of change is greater than 3) are extracted as independent copper loss degradation samples. In the iron loss-dominated mode, the z-score normalization method is also used to process the iron loss value measurement data series to calculate its rate of change. The periods in which iron loss shows a continuous growth trend and the growth rate exceeds the preset iron loss growth rate threshold (e.g., the growth rate is greater than 1.5 for more than 24 hours) are specially marked as "potential associated degradation source" periods.

[0038] z-score standardization unifies loss data of different dimensions and amplitudes to a standard normal distribution, facilitating the identification of abnormal fluctuations through set thresholds. For example, if a transformer's copper loss suddenly and sharply increases on a certain day, with its normalized rate of change reaching 4.1, far exceeding the copper loss rate of change threshold of 3, this may indicate an inter-turn short circuit in the winding, and the system marks this as "independent copper loss degradation." In another period, its iron loss increases slowly but steadily for three consecutive days under light load mode, with a normalized growth rate maintained at around 1.8, exceeding the iron loss growth rate threshold of 1.5. The system marks this as a "potential associated degradation source" because this continuous deterioration of iron loss (possibly caused by aging of the core silicon steel sheets or loosening of clamping force) is the fundamental driving force behind the "thermal bridge" effect. This step, through separation analysis, preliminarily distinguishes between two scenarios: one potentially caused by the winding itself, and the other potentially caused by the core. As an alternative, in addition to z-score, min-max normalization can also be used; for anomaly detection, algorithms based on Isolation Forest can be used, which are more effective in processing high-dimensional data and identifying novel anomalies.

[0039] Step S4 involves a quantitative assessment of the cross-domain correlation between iron loss and copper loss. This step is crucial for achieving the objectives of this invention. It initiates in-depth cross-domain correlation analysis on the "potential correlation degradation source" time period marked in step S3 to quantitatively assess the driving force of iron loss anomalies on copper loss.

[0040] Reference Figure 2 The figure illustrates the core principle of this step. After identifying potential associated degradation source periods, the system extracts the time series of iron loss and copper loss change rates within those periods. Subsequently, the cross-correlation function between these two time series is calculated, which can be expressed as R_xy(τ) = E[(X_t - μ_x)(Y_{t+τ} - μ_y)]. In this formula, X_t and Y_t are the iron loss and copper loss change rate series, respectively; μ_x and μ_y are the means of series X_t and Y_t, respectively; τ is the time delay; and E represents the expected value. The calculated result is a function of the time delay τ. The magnitude of the peak value of this function represents the maximum correlation strength between the two series at a certain time delay, and the horizontal axis corresponding to the peak value is the maximum delay time τ_peak. This invention defines the peak value of this cross-correlation function as the Cross-domain Influence Factor (CDIF). The time lag τ at the peak of this function represents the time required for iron loss changes to have the greatest impact on copper loss changes, while the magnitude of the peak itself represents the strength of this impact. A CDIF value significantly higher than the normal fluctuation range indicates a strong cross-domain correlation, meaning that the deterioration of iron loss is dominating the abnormal growth of copper loss.

[0041] Step S5 involves performing degradation prediction modeling based on fused correlation features. This step constructs a prediction model capable of identifying multiple failure modes. Specifically, a training sample set fusing multi-dimensional features is constructed, where each sample contains: the standardized loss trajectory for the corresponding time period (from step S3), the corresponding operating condition parameter input sequence (from step S2), and the correlation quantification index calculated in step S4 (CDIF or scalar features extracted from CDIM). A support vector machine (SVM) algorithm is used to construct the fused degradation prediction model. By training with feature vectors containing CDIF or CDIM features, the SVM model can learn and identify the "iron loss-copper loss cross-domain correlation degradation" mode characterized by highly correlated feature values.

[0042] Reference Figure 4 This figure further illustrates the role of the CDIF feature proposed in this invention in improving the prediction accuracy of the model. (a) shows the distribution of healthy samples (represented by hollow circles) and associated degraded samples (represented by crosses) in the traditional feature space (such as normalized iron loss and copper loss change rates). It can be seen that there is a significant overlap between the two types of samples, making it difficult for a linear classifier (as shown by the dashed line in the figure, the SVM decision boundary) to effectively separate them. (b) shows the feature space fused with the CDIF of this invention, with one coordinate axis replaced by the CDIF proposed in this invention, forming a new feature space. In this space, the CDIF values ​​of healthy samples are concentrated in the low range (e.g., 0-0.2), while the CDIF values ​​of associated degraded samples are significantly concentrated in the high range (e.g., 0.6-0.9), forming a clear boundary. This allows the SVM decision boundary to easily separate the two types of samples completely. This intuitively demonstrates that CDIF, as a key feature, can significantly enhance the separability of samples, thus laying the foundation for building a high-precision degradation prediction model.

[0043] In practical implementation, the feature vector can be constructed in the form of [z_cu_rate (normalized copper loss rate of change), z_fe_rate (normalized iron loss rate of change), I_avg, T_avg, V_avg, CDIF], where I_avg is the average load current during the corresponding analysis period, T_avg is the average operating temperature (e.g., winding temperature or core temperature) during that period, and V_avg is the average input voltage during that period. A large number of historical data samples with labels (e.g., "healthy," "independent degradation," "correlated degradation") are input into the SVM classifier for training. During training, the SVM maps the data to a high-dimensional feature space by selecting an appropriate kernel function (e.g., the radial basis function (RBF) kernel), and finds an optimal hyperplane in this space that can separate samples of different classes with the maximum margin. Because the feature vector contains the key physical feature CDIF, the trained model can capture fault modes that are ignored by traditional methods. For example, when a sample's z_cu_rate and z_fe_rate (representing the normalized rates of change in copper loss and iron loss, respectively) are not extremely abnormal, but its CDIF value is high, traditional models may misclassify it as healthy, while the model of this invention can correctly identify it as an early stage of "associated degradation." Using CDIM, its determinant or maximum eigenvalue can be extracted as input features. These scalars integrate multidimensional information from the matrix, making the model more robust to understanding the degradation mechanism. As an alternative, deep learning models, such as Long Short-Term Memory (LSTM) networks, can be used. These networks can automatically learn features and temporal dependencies from the original time series, potentially saving some manual feature engineering, but requiring more data and computational resources.

[0044] Step S6 involves tiered early warning and risk tracing. This step interprets and applies the model's prediction results. First, a preset threshold judgment mechanism is used to classify the degradation tendency probability values ​​output by the model in Step S5 into risk levels. For example, if the output probability is greater than 0.9, it is judged as high degradation risk; if the probability is between 0.7 and 0.9, it is judged as medium degradation risk. When the judgment result is high or medium risk, an early warning is triggered. At the same time, in order to achieve accurate risk tracing, the feature vector at the current moment is obtained, and the cosine similarity between it and the feature vector of the high degradation risk samples already labeled in the historical database is calculated. If the calculated similarity exceeds a preset similarity threshold (e.g., 0.95), not only is a degradation early warning signal triggered, but also a refined identification result is generated based on the label of the most similar historical high-risk sample, clearly indicating whether it is "independent copper loss degradation" or "cross-domain associated degradation of iron loss and copper loss".

[0045] Reference Figure 6This figure illustrates the complete monitoring and early warning process of the method of this invention with a specific example. (a) shows the original loss data trajectory, displaying the original data trajectories of iron loss (solid line) and copper loss (dashed line) within a 60-day monitoring period. It can be seen that iron loss begins to slowly and abnormally increase around day 20, while copper loss reaches an inflection point around day 35, with a significant time delay τ between the two. (b) shows the evolution of the Cross-Domain Influence Factor (CDIF). After the iron loss anomaly, the CDIF value begins to rise continuously, significantly exceeding the preset early warning threshold (e.g., 0.5) before the copper loss anomaly. (c) shows the output risk level early warning, displaying the risk level ladder diagram generated by the system based on the model output. This achieves a step-by-step early warning from low risk (white area) to medium risk (diagonally filled area) and then to high risk (cross-filled area), providing clear and intuitive decision support for maintenance personnel.

[0046] Reference Figure 7 The figure further illustrates the technical principles of risk tracing and pattern recognition in step S6. In a two-dimensional feature space, two types of historical fault samples are distributed: "independent copper loss degradation" (represented by triangles) and "cross-domain associated degradation of iron loss and copper loss" (represented by squares). When a current test sample (represented by a star) appears, the system calculates the cosine similarity between its feature vector and the feature vectors of all historical samples in the database. As shown in the figure, the vector angle between the test sample and sample P (a known "cross-domain associated degradation of iron loss and copper loss" sample) is the smallest, i.e., the cosine similarity is the highest (e.g., reaching 0.98, exceeding the threshold of 0.95). Based on this, the system can accurately identify the current test sample as a "cross-domain associated degradation of iron loss and copper loss" pattern, thereby achieving accurate risk tracing.

[0047] In a preferred embodiment, during the dynamic identification of the operating mode in step S2, if the load current monitoring value continuously exceeds 80% of the transformer's rated current within a certain time period, then that time period is determined to be the copper loss-dominated mode, because the load current is large at this time, and copper loss plays an absolutely dominant role in the total loss; if the load current monitoring value continuously falls below 30% of the rated current, then that time period is determined to be the iron loss-dominated mode, because the load current is small at this time, and iron loss, which is unrelated to the load, becomes the main part of the total loss.

[0048] Furthermore, to more comprehensively analyze the complex coupling relationship between iron loss and copper loss, in a preferred embodiment of the present invention, the aforementioned single, one-dimensional cross-domain influence factor is extended into a multi-dimensional "Cross-domain Influence Matrix (CDIM)". This matrix is ​​constructed by parallel computation of cross-correlation functions between multiple sets of different physical quantities, and can capture the correlation characteristics between the two from different dimensions. For example, the CDIM can be a 2x2 matrix, and its elements are defined as follows:

[0049] Matrix element (1,1): The cross-correlation peak of the iron loss rate change sequence and the copper loss rate change sequence, i.e. the original CDIF, which characterizes the synchronicity of the changing trends of the two.

[0050] Matrix element (1,2): The cross-correlation peak of the absolute value sequence of iron loss and the rate of change sequence of copper loss, used to characterize the driving strength of the cumulative effect of iron loss (such as the continuous high temperature of the iron core) on the instantaneous change of copper loss.

[0051] Matrix element (2,1): The cross-correlation peak of the iron loss rate sequence and the copper loss absolute value sequence, used to characterize the impact of instantaneous drastic changes in iron loss (such as a surge in instantaneous iron loss due to overvoltage) on the cumulative level of copper loss.

[0052] Matrix element (2,2): The cross-correlation peak of the iron loss change acceleration sequence and the copper loss change acceleration sequence, used to characterize the synchronicity of the drastic changes of the two, and can be used to capture the acceleration phase of degradation.

[0053] Reference Figure 5 This figure visually illustrates the performance of CDIM under different health states using a heatmap format. (a) shows CDIM in a healthy state, where matrix element values ​​are generally low (lighter color), reflecting a weak correlation between iron loss and copper loss. (b) shows CDIM in a degraded state, where both diagonal and off-diagonal elements exhibit significantly high values ​​(darker color), demonstrating a multi-dimensional and strongly coupled correlation between the two. By constructing such a cross-domain influence matrix, it is possible to expand from "point correlation" to "area correlation," depicting various complex coupling modes between iron loss and copper loss, including long-term slow effects and short-term severe shocks.

[0054] As a preferred implementation, when performing degradation prediction modeling of fused associated features in step S5, a Support Vector Machine (SVM) can be used as the machine learning algorithm. SVM is a powerful classification algorithm that maps input features to a high-dimensional space through a kernel function (such as a radial basis function kernel) and finds an optimal hyperplane in that space to separate samples of different categories with the maximum margin.

[0055] In the preferred embodiment described above, when using the Cross-Domain Influence Matrix (CDIM) for association quantification, to integrate matrix information into conventional classifiers such as SVM, instead of directly using a single CDIF as input features, one or more key scalars extracted from the CDIM are used as more robust and informative core features. For example, the determinant, trace (sum of the main diagonal elements), or maximum eigenvalue of the CDIM can be calculated and used. The determinant of the matrix can comprehensively reflect the overall strength of multi-dimensional associations, the trace of the matrix can represent the sum of the main association effects, and the maximum eigenvalue can capture the most significant association patterns. Inputting these more informative scalar features into the SVM model enables the model to analyze degradation mechanisms, thereby improving the accuracy and robustness of predictions.

[0056] As a preferred implementation, during step S6 for graded early warning and risk tracing, a preset threshold judgment mechanism can be set to classify the risk level of the degradation tendency probability value output by the model. For example, if the output probability is greater than 0.9, it is judged as high degradation risk; if the probability is between 0.7 and 0.9, it is medium degradation risk; and if the probability is less than 0.7, it is low degradation risk. During risk tracing, cosine similarity can be used to measure the similarity between the current feature vector and the feature vectors of historical high-risk samples. This similarity is measured by calculating the cosine value of the angle between the two vectors; the closer the value is to 1, the more similar the degradation patterns of the two samples are. If the calculated similarity exceeds a preset similarity threshold (e.g., 0.95), not only is a degradation early warning signal triggered, but also a refined identification result is generated based on the label of the most similar historical high-risk sample. This result clearly includes the risk level, similarity value, and specific degradation pattern type.

[0057] Example 2

[0058] This invention also provides an intelligent monitoring device for transformer operating energy efficiency, the structural block diagram of which is shown below. Figure 3 As shown. The device can be a hardware entity integrated into a transformer monitoring terminal, or a software program deployed on a cloud server or edge computing device. The device includes:

[0059] A data acquisition module 10 is used to execute step S1 in the aforementioned method embodiment. This module connects to the sensors or database of the transformer online monitoring system through a communication interface, acquires data such as iron loss value, copper loss value, load current, operating temperature and voltage level at a fixed sampling frequency (e.g., every 5 minutes), and timestamps the data to construct a multi-dimensional time-series dataset.

[0060] An operating mode recognition module 20 is used to execute step S2 in the aforementioned method embodiment. This module receives the dataset output by the data acquisition module 10, compares the load current monitoring value with preset thresholds (such as 30% and 80% of the rated current), determines the current operating mode of the transformer (iron loss dominant or copper loss dominant), and outputs a data stream with a mode identifier.

[0061] An independent loss analysis module 30 is used to execute step S3 in the aforementioned method embodiment. This module performs z-score normalization on the corresponding dominant loss data sequence based on the pattern identifier output by the operation mode recognition module 20, calculates its rate of change, and filters out independent copper loss degradation samples according to the preset first and iron loss growth rate thresholds, while marking the time period of "potential associated degradation source".

[0062] A cross-domain correlation assessment module 40 is used to perform step S4 in the aforementioned method embodiment. When a signal of a "potentially correlated degradation source" time period marked by the independent loss analysis module 30 is received, it starts a calculation program to extract the correlation time series of iron loss and copper loss within that time period, calculate the cross-correlation function between the two, and finally generate features for quantifying the strength of cross-domain correlation, such as the cross-domain influence factor (CDIF) or the cross-domain influence matrix (CDIM).

[0063] A predictive modeling module 50 is used to execute step S5 in the aforementioned method embodiment. This module internally embeds or dynamically loads a pre-trained fusion degradation prediction model, such as a support vector machine (SVM) model. It receives standardized loss trajectories, operating parameters, and correlation features generated by the cross-domain correlation evaluation module 40 from various modules, combines them into a complete feature vector, inputs it into the prediction model, and finally outputs a probability value representing the degradation tendency.

[0064] An early warning and source tracing module 60 is used to execute step S6 in the aforementioned method embodiment. This module receives the probability value output by the predictive modeling module 50 and judges it according to preset risk level thresholds (such as 0.7 and 0.9) to trigger early warning signals of different levels. At the same time, this module is also connected to a historical fault case database. When an early warning is triggered, it calculates the cosine similarity between the current state feature vector and the samples in the database to achieve accurate source tracing of the fault mode and generate a refined identification result report containing risk level, similarity, and specific degradation type.

[0065] It is understood that the functions of each module in the above-mentioned device can be implemented through a central processing unit (CPU), digital signal processor (DSP), or application-specific integrated circuit (ASIC) in conjunction with corresponding software instructions. The modules communicate and exchange data with each other through an internal bus or data interface, working collaboratively to complete the intelligent monitoring method proposed in this invention.

[0066] Example 3

[0067] To further illustrate the technical effects of the present invention, a specific application example is given below as Example 3.

[0068] This embodiment applies to an SCB13-2500kVA / 10kV / 0.4kV dry-type epoxy resin cast transformer installed in a large data center. This transformer provides highly reliable and continuous power to a large number of server racks within the data center. During the initial deployment phase, the system continuously collects and stores multi-dimensional time-series data, including iron loss power (P_fe), copper loss power (P_cu), three-phase load current (I_A, I_B, I_C), winding hot spot temperature (T_w), and core surface temperature (T_c), with a sampling period of 5 minutes. A health status baseline model of the transformer is established using stable operating data from the first three months.

[0069] When the system reached its fourth month of operation, the independent loss trajectory analysis module (corresponding to step S3) first identified, through the dynamic identification module for operating modes (corresponding to step S2), a low-load period from 2:00 AM to 4:00 AM daily as server maintenance, with load current consistently below 30% of the rated current. This is a typical "iron loss-dominated mode." Under this mode, the analysis module detected a continuous, slight upward trend in iron loss values. After z-score normalization, the rate of change remained stable between 1.6 and 1.9 for several consecutive days. Although the value was small, it exceeded the system's preset iron loss growth rate threshold of 1.5. Based on this, the system automatically marked this period as a "potential associated degradation source" and upgraded the monitoring level for the transformer. However, no external warning was triggered at this time because other indicators such as copper loss were within the normal range.

[0070] After marking the "potential associated degradation source" period, the system automatically triggered the core algorithm for quantitative assessment of the cross-domain correlation between iron loss and copper loss (corresponding to step S4). The system extracted the time series of iron loss change rate and copper loss change rate over the past 60 days and calculated the cross-correlation function between them. The calculation results revealed a key phenomenon: at a time delay τ of 48 hours, the cross-correlation function showed a significant peak of 0.85. This peak was quantified as the Cross-Domain Influence Factor (CDIF) = 0.85. This result physically reflects that the high-order harmonics generated by the nonlinear load of the data center (server switching power supply) cause the iron loss to continuously deteriorate. The additional heat generated forms a "thermal bridge" through the compact solid structure. After about two days of heat accumulation and conduction, it begins to significantly affect the resistivity of the winding, thus causing an abnormal synchronous increase in copper loss. This time delay and intensity from iron loss deterioration to copper loss response is successfully captured by the CDIF parameter. Furthermore, the system also constructed a cross-domain influence matrix (CDIM), whose determinant value deviated significantly from the baseline range under healthy conditions, confirming the strength and complexity of this association from multiple dimensions.

[0071] Subsequently, the degradation prediction modeling module that integrates the associated features (corresponding to step S5) inputs a multidimensional feature vector, including this high CDIF value (0.85) and the determinant scalar derived from CDIM, into the pre-trained support vector machine (SVM) model. Although the absolute value and rate of change of copper loss have not yet triggered independent anomaly thresholds at this point, due to the significant increase in the key feature CDIF, the SVM model, after comprehensive judgment, outputs a degradation tendency probability value of 0.96 for the "iron loss-copper loss cross-domain associated degradation" pattern.

[0072] Based on this high-probability output, the early warning and source tracing module (corresponding to step S6) immediately judges this probability as a high degradation risk and triggers a level-one early warning. The early warning information does not simply report "abnormal loss," but performs in-depth source tracing. The module performs cosine similarity matching between the current feature vector and the historical fault case library, finding a similarity of 0.97 with case number "F-DT-018" (the fault cause being "excessive harmonic current in the server cluster leading to localized overheating of the iron core, thus accelerating winding insulation aging"). The system ultimately generates a diagnostic report, including: High-risk warning (degradation tendency probability value 96%): Detected a "cross-domain correlation degradation of iron loss and copper loss." The fault characteristics are highly similar to historical case F-DT-018 (harmonics causing iron core overheating) (similarity 97%).

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent monitoring of transformer operating energy efficiency, characterized in that, Includes the following steps: Obtain the measurement data of iron loss value, copper loss value, and load current monitoring value of the transformer, and construct the iron loss time series and copper loss time series; Based on the transformer's load current monitoring values, the periods when the load current is lower than the preset low load threshold and the iron loss-dominated mode is identified. During the period of the iron loss-dominant mode, the iron loss value measurement data sequence is normalized and the rate of change of iron loss is calculated; Periods in which the rate of change of iron loss continuously exceeds a preset threshold for the rate of increase of iron loss are marked as potential associated degradation source periods in order to identify potential associated degradation source periods in which iron loss shows a continuous increasing trend. For the time period of the potential associated degradation source, the cross-correlation function between the time series of iron loss change rate and the time series of copper loss change rate is calculated, and the peak value of the cross-correlation function is used as the cross-domain influence factor to quantify the cross-domain correlation strength between iron loss and copper loss. Based on the aforementioned cross-domain impact factors, the risk of cross-domain correlation degradation of iron loss and copper loss in transformers is predicted. The prediction process includes the following degradation pattern identification sub-steps: First, construct a current state feature vector containing the normalized copper loss rate of change, the normalized iron loss rate of change, the average load current, the average operating temperature, the average input voltage, and the cross-domain influence factor. Secondly, the similarity of the current state feature vector with the feature vectors of multiple historical samples stored in the historical database that have been marked as different degradation modes is calculated one by one to obtain a set of similarity values ​​that correspond one-to-one with the historical samples. Then, a maximum similarity value is determined from this set of similarity values, and the degradation pattern label of the historical sample corresponding to the maximum similarity value is obtained; Finally, when the maximum similarity value exceeds the preset similarity threshold, the obtained degradation mode label is output as the final identification result. The content of the identification result is a label indicating the specific degradation mode type of the current transformer. The degradation mode type is copper loss independent degradation or iron loss-copper loss cross-domain associated degradation.

2. The method according to claim 1, characterized in that, The cross-domain influence factor is extended into a cross-domain influence matrix, which is constructed by parallel calculation of the peak values ​​of the cross-correlation functions between multiple sets of different physical quantity time series. The multiple sets of different physical quantity time series are selected from at least two sets of iron loss absolute value series, copper loss absolute value series, iron loss change rate time series, copper loss change rate time series, iron loss change acceleration series, and copper loss change acceleration series.

3. The method according to claim 2, characterized in that, Also includes: Extract the determinant, trace, or maximum eigenvalue of the cross-domain influence matrix as features to quantify the strength of cross-domain associations. Based on the extracted determinant, matrix trace, or maximum eigenvalue, the risk of cross-domain degradation of iron loss and copper loss in the transformer is predicted.

4. The method according to claim 3, characterized in that, The steps for predicting whether a transformer has a cross-domain correlation degradation risk between iron loss and copper loss based on the extracted determinant, matrix trace, or maximum eigenvalue include: Obtain the input sequence of operating parameters reflecting the transformer's operating conditions; Construct a feature vector that integrates features extracted from the cross-domain influence matrix to quantify the cross-domain correlation strength and the input sequence of the operating condition parameters; The feature vector is input into a pre-trained degradation prediction model, which outputs a degradation tendency probability value that represents the degradation tendency.

5. The method according to claim 4, characterized in that, The degradation prediction model is a support vector machine model, and the method further includes: mapping the feature vector to a high-dimensional feature space using a radial basis function kernel, and determining an optimal hyperplane in the high-dimensional feature space to separate samples of different degradation modes with the maximum interval, so as to construct the support vector machine model.

6. The method according to claim 4, characterized in that, It also includes a graded early warning step, which includes: comparing the degradation tendency probability value with a high-risk threshold used to distinguish between medium and high risk levels and a medium-risk threshold used to distinguish between low and medium risk levels, so as to classify the current transformer risk level into one of the three levels of high, medium and low, and triggering the corresponding early warning signal according to the risk level.

7. The method according to claim 1, characterized in that, The similarity is the cosine of the angle between the current state feature vector and the feature vector of the historical sample.

8. The method according to claim 1, characterized in that, In the process of acquiring transformer iron loss measurement data, copper loss measurement data, and load current monitoring data, and constructing iron loss time series and copper loss time series, load current monitoring data, operating temperature data, and voltage level recording data are collected simultaneously, and timestamps are added to the collected data of each dimension to construct a multi-dimensional time series structure dataset.

9. A transformer operating energy efficiency intelligent monitoring device, characterized in that, include: A data acquisition module is used to acquire the measurement data of iron loss value, copper loss value and load current monitoring value of transformer, and construct iron loss time series and copper loss time series; An independent loss analysis module is used to identify periods when the load current is below a preset low load threshold, based on the transformer's load current monitoring values, that are dominated by iron loss. During the period of the iron loss-dominant mode, the iron loss value measurement data sequence is normalized and the rate of change of iron loss is calculated; Periods in which the rate of change of iron loss continuously exceeds a preset threshold for the rate of increase of iron loss are marked as potential associated degradation source periods in order to identify potential associated degradation source periods in which iron loss shows a continuous increasing trend. A cross-domain correlation assessment module is used to calculate the cross-correlation function between the time series of iron loss change rate and the time series of copper loss change rate for the potential correlation degradation source period, and use the peak value of the cross-correlation function as a cross-domain influence factor to quantify the cross-domain correlation strength between iron loss and copper loss. A predictive modeling module is used to predict whether there is a risk of cross-domain correlation degradation between iron loss and copper loss in transformers based on the cross-domain influencing factors. A degradation mode identification module is configured to perform the following steps: First, construct a current state feature vector containing normalized copper loss change rate, normalized iron loss change rate, average load current, average operating temperature, average input voltage, and the cross-domain influence factor; second, perform similarity calculations on the current state feature vector and the feature vectors of multiple historical samples labeled with different degradation modes stored in a historical database to obtain a set of similarity values ​​corresponding one-to-one with the historical samples; then, determine a maximum similarity value in the set of similarity values ​​and obtain the degradation mode label of the historical sample corresponding to the maximum similarity value; finally, when the maximum similarity value exceeds a preset similarity threshold, output the obtained degradation mode label as the final identification result, wherein the identification result is a label indicating the specific degradation mode type occurring in the current transformer, and the degradation mode type is either independent copper loss degradation or cross-domain associated degradation of iron loss and copper loss.

Images