A power transformer fault early warning method and system

Abstract

This invention relates to the field of power equipment monitoring and fault diagnosis technology, and discloses a method and system for early warning of power transformer faults. The method includes collecting the transformer's load current, ambient temperature, and top oil temperature sequences; calculating the initial measured temperature rise based on the top oil temperature and ambient temperature, and determining the thermal response lag time through cross-correlation analysis; shifting and correcting the load current according to the lag time, and constructing a time-aligned excitation input set and measured temperature rise sequence based on a common effective time window; substituting the excitation input set into a thermal circuit model based on the thermal balance differential equation to calculate the theoretical reference temperature rise, combining the measured temperature rise sequence to calculate the residual and performing moving average filtering to construct a fault feature vector; inputting the fault feature vector into a fault diagnosis classification model to determine the fault type and severity, and generating a graded early warning command accordingly. This method can eliminate thermal lag interference and achieve accurate quantitative early warning of latent thermal faults in transformers under dynamic operating conditions.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring and fault diagnosis technology, and in particular to a method and system for early warning of power transformer faults. Background Technology

[0002] Currently, power transformers, as the core hub for energy transmission and conversion in the power grid, directly affect the safety and stability of the entire power system through their operational reliability. With the advancement of smart grid construction, the maintenance strategy for power equipment is gradually shifting from traditional "reactive maintenance" and "periodic inspection" to "predictive maintenance" based on condition awareness. This has made prognostics and health management (PHM) technology for key transformer components a current research hotspot and focus in the industry.

[0003] In existing technologies, single-threshold monitoring methods based on oil surface thermometers or offline diagnostic methods based on dissolved gas analysis (DGA) are commonly used to assess the operating status of transformers. These methods mainly involve setting fixed upper limits for temperature or gas content thresholds, triggering alarms when monitored data exceeds preset standards, or determining whether there are abnormalities in the equipment based on static indicators at a single time point. Because a transformer is a physical system with significant thermal inertia, its internal temperature rise exhibits a significant thermal response lag relative to changes in load current, and it is highly susceptible to sudden changes in ambient temperature. Traditional static thresholds or simple models often ignore this dynamic thermal conduction lag characteristic and its coupling effect with the environment, leading to false alarms (treating normal thermal delays as faults) during severe load fluctuations or sudden environmental changes, and missed detections when early, weak faults occur due to being masked by ambient noise. This fails to meet the real-time and robust requirements of high-precision fault prediction and health management.

[0004] Existing technologies have the problem of accurately distinguishing between normal thermal fluctuations and early latent faults under dynamic and complex operating conditions. Summary of the Invention

[0005] This invention provides a method and system for early warning of power transformer faults, in order to solve the problem in the prior art that it is difficult to accurately distinguish between normal thermal fluctuations and early latent faults under dynamic and complex operating conditions.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a power transformer fault early warning method, comprising:

[0007] Collect real-time operating status data of the transformer, including load current sequence, ambient temperature sequence and top oil temperature sequence;

[0008] The initial measured temperature rise sequence is obtained by performing a difference calculation between the top oil temperature sequence and the ambient temperature sequence, and a cross-correlation analysis is performed between the load current sequence and the initial measured temperature rise sequence to determine the thermal response hysteresis time.

[0009] The load current sequence is time-shifted according to the thermal response hysteresis time to determine the common effective time window of the shifted load current sequence, the ambient temperature sequence, and the initial measured temperature rise sequence. Synchronous data is extracted based on the common effective time window to construct the time-aligned excitation input set and the time-aligned measured temperature rise sequence.

[0010] The time-aligned excitation input set is substituted into a preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence.

[0011] The time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence are subjected to time-domain difference operation to obtain the temperature rise residual sequence. The temperature rise residual sequence is then subjected to moving average filtering. A fault feature vector is constructed based on the processed temperature rise residual sequence.

[0012] The fault feature vector is input into a preset fault diagnosis classification model to determine the potential fault type label and fault severity level;

[0013] If the severity level of the fault exceeds a preset safety threshold, a corresponding graded warning instruction is generated based on the potential fault type label.

[0014] Secondly, the present invention provides a power transformer fault early warning system, comprising:

[0015] The data acquisition module is used to acquire real-time operating status data of the transformer, including load current sequence, ambient temperature sequence and top oil temperature sequence.

[0016] The time series analysis module is used to perform difference calculations based on the top oil temperature sequence and the ambient temperature sequence to obtain the initial measured temperature rise sequence, and to perform cross-correlation analysis on the load current sequence and the initial measured temperature rise sequence to determine the thermal response hysteresis time.

[0017] The data alignment module is used to perform time-series shifting of the load current sequence based on the thermal response hysteresis time, determine the common effective time window of the shifted load current sequence, the ambient temperature sequence, and the initial measured temperature rise sequence, and perform synchronous data interception based on the common effective time window to construct the time-aligned excitation input set and the time-aligned measured temperature rise sequence.

[0018] The model calculation module is used to substitute the time-aligned excitation input set into a preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence.

[0019] The feature extraction module is used to perform time-domain difference operation on the time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence to obtain a temperature rise residual sequence, and to perform moving average filtering on the temperature rise residual sequence to construct a fault feature vector based on the processed temperature rise residual sequence.

[0020] The diagnostic decision module is used to input the fault feature vector into a preset fault diagnosis classification model to determine the potential fault type label and fault severity level;

[0021] The early warning response module is used to generate a corresponding graded early warning instruction based on the potential fault type label if the severity level of the fault exceeds a preset safety threshold.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) This invention accurately calculates the thermal response lag time between the load and the temperature rise by introducing cross-correlation analysis technology, and performs time-series shift correction and common window truncation on the load current sequence accordingly. This mechanism effectively overcomes the "temperature rise delay" phenomenon of transformers as large thermal inertia systems when the load fluctuates drastically, and solves the problem of residual miscalculation caused by the asynchronous dynamic and static models in traditional static models; through physical time alignment, it ensures strict synchronization between the excitation input and the thermal response in the time domain, and significantly improves the robustness and accuracy of fault prediction under dynamic operating conditions.

[0024] (2) This invention generates a theoretical reference temperature rise by constructing a transformer thermal circuit model based on the thermal balance differential equation, and performs a differential operation between the time-aligned measured temperature rise and the theoretical reference to obtain the temperature rise residual sequence. This method effectively decouples the ambient temperature floor effect and transforms complex absolute temperature monitoring into pure fault residual analysis. Compared with traditional absolute threshold monitoring, this invention can extract weak early fault feature signals from strong ambient noise, greatly improving the detection sensitivity of latent thermal faults such as winding micro-short circuits and poor contact.

[0025] (3) This invention performs moving average filtering and feature vector construction on the temperature rise residual sequence, and combines it with a fault diagnosis classification model to assess the severity. This process not only effectively filters out high-frequency random noise during sensor acquisition, but also realizes the transformation from single numerical alarm to fault severity quantification; the differentiated graded early warning instructions generated according to the risk level can assist maintenance personnel in taking targeted measures before the fault evolves into a catastrophic accident, providing a scientific and objective decision-making basis for fault prediction and health management throughout the entire life cycle of power transformers. Attached Figure Description

[0026] Figure 1 This is a schematic flowchart of a power transformer fault early warning method provided in the first embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of a power transformer fault early warning system provided in the second embodiment of the present invention. Detailed Implementation

[0028] 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 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.

[0029] Reference Figure 1 The first embodiment of the present invention provides a power transformer fault early warning method, including the following steps:

[0030] S11, Collect real-time operating status data of the transformer, including load current sequence, ambient temperature sequence and top oil temperature sequence;

[0031] S12, Based on the difference calculation between the top oil temperature sequence and the ambient temperature sequence, an initial measured temperature rise sequence is obtained, and cross-correlation analysis is performed between the load current sequence and the initial measured temperature rise sequence to determine the thermal response hysteresis time;

[0032] S13, the load current sequence is time-shifted according to the thermal response hysteresis time, the common effective time window of the shifted load current sequence, the ambient temperature sequence and the initial measured temperature rise sequence is determined, and synchronous data is extracted based on the common effective time window to construct the time-aligned excitation input set and the time-aligned measured temperature rise sequence.

[0033] S14, Substitute the time-aligned excitation input set into the preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence.

[0034] S15, perform time-domain difference operation on the time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence to obtain the temperature rise residual sequence, and perform moving average filtering on the temperature rise residual sequence to construct a fault feature vector based on the processed temperature rise residual sequence.

[0035] S16, Input the fault feature vector into a preset fault diagnosis classification model to determine the potential fault type label and fault severity level;

[0036] S17. If the severity level of the fault exceeds a preset safety threshold, a corresponding graded warning instruction is generated based on the potential fault type label.

[0037] In step S11, real-time operating status data of the transformer is collected. This real-time operating status data includes load current sequence, ambient temperature sequence, and top oil temperature sequence, including:

[0038] The transformer's top oil temperature, ambient temperature, and load current are continuously and synchronously collected using temperature sensors and current transformers at a preset sampling frequency.

[0039] The collected continuous numerical points are arranged in order of collection timestamp to generate a raw data stream containing the time dimension;

[0040] The original data stream is preprocessed by imputing missing values ​​and removing outliers, and the load current sequence, the ambient temperature sequence, and the top oil temperature sequence are extracted by column.

[0041] It should be noted that the continuous synchronous data acquisition via temperature sensors and current transformers is achieved using a distributed data acquisition architecture based on the Network Time Protocol (NTP). This architecture consists of platinum resistance thermometers (such as PT100) deployed on top of the transformer tank, ambient temperature and humidity sensors deployed around the transformer, and a signal conditioning module connected to the secondary side of the current transformer (CT) connected to the transformer bushing. Under the triggering of a unified clock source, each sensor converts the physical analog quantities into digital signals and adds a timestamp accurate to the millisecond level, thereby ensuring strict alignment of the three sets of data in the time dimension.

[0042] It is worth noting that the preset sampling frequency is determined based on the Nyquist-Shannon Sampling Theorem and statistical analysis of the transformer's thermal inertia characteristics. The system first obtains the minimum thermal time constant of the transformer oil temperature change. (Measured by a transformer temperature rise test, for example, over 2 hours). To accurately reconstruct the dynamic details of temperature rise changes while avoiding data redundancy, the sampling interval... Should meet ,in This is a preset oversampling factor. In this embodiment, based on statistics of temperature rise rates in historical fault cases, the factor is selected. The calculated sampling frequency is once per minute (i.e., 1 / 60Hz) to capture minute-level temperature rise abrupt changes. The preset oversampling factor... The setting is based on the ratio of the transient frequency of temperature rise induced by latent faults in transformers to the thermal equilibrium response time. To ensure the capture of nonlinear abrupt changes during the thermal equilibrium transition, at least 100 sampling points must exist within a thermal time constant. In this embodiment, based on statistics of temperature rise rates in historical fault cases, [the following is selected]. The sampling frequency was calculated to be once per minute (i.e., 1 / 60Hz) to capture minute-level temperature rise abrupt changes.

[0043] It should be noted that the preprocessing of the original data stream, including missing value imputation and outlier removal, is performed using linear interpolation and a sliding window method. The criteria are jointly implemented. First, for missing values ​​(NaN) imputation, if the number of consecutive missing points is less than a preset missing value threshold... (For example, 3), then utilize the valid data points before and after the missing points. and Construct the linear interpolation equation:

[0044]

[0045] The fill value was calculated. The missing interruption threshold The setting is based on the quasi-steady-state assumption of the transformer thermal process, that is, within a very short time (e.g., 3 minutes), the oil temperature change can be considered linear. If the missing time exceeds this threshold, the linear interpolation is considered invalid and needs to be marked as invalid data segment to prevent the introduction of artificial noise. Secondly, for the removal of outliers, a length of [missing information] is constructed. Sliding time window (e.g.) The length of the sliding window. The setting is based on the coupling period of the transformer load periodic fluctuation characteristics and the daily variation rate of ambient temperature; [selection / selection] (Corresponding to 30 minutes) This ensures that the window contains sufficient locally stationary samples to accurately estimate the statistical distribution parameters under the current operating conditions, while avoiding interference from trend terms caused by excessively large spans. Calculate the mean of the data within the window. and standard deviation If the current data point satisfy If the value is an outlier, it is identified as an extreme value and removed. The resulting value is then corrected and replaced using the interpolation equation described above. This operation ensures the numerical continuity and physical plausibility of the subsequently extracted sequences.

[0046] For example, the sampling frequency is set to 1 minute / time. In the collected raw top-layer oil temperature data stream, the temperature is 75.0℃ at 10:00, 75.4℃ at 10:02, but data at 10:01 is lost. The system calculates the value at 10:01 using linear interpolation. ℃. If the value collected at 10:03 suddenly changes to 150.0℃ (due to strong interference), while the mean of the current sliding window (past 30 minutes) is 75.1℃ and the standard deviation is 0.2℃. Because The system determined that 150.0℃ was an outlier and corrected it to the interpolation of a nearby point (such as 75.6℃). Finally, the load current sequence, ambient temperature sequence and top oil temperature sequence after cleaning were extracted respectively.

[0047] In step S12, an initial measured temperature rise sequence is obtained by performing a difference calculation based on the top oil temperature sequence and the ambient temperature sequence, and a cross-correlation analysis is performed on the load current sequence and the initial measured temperature rise sequence to determine the thermal response hysteresis time, including:

[0048] Subtract the corresponding value in the ambient temperature sequence from the value at each moment in the top oil temperature sequence to generate the initial measured temperature rise sequence;

[0049] Set a time sliding window and calculate the cross-correlation function between the load current sequence and the initial measured temperature rise sequence at different time offsets;

[0050] Search for the maximum peak point of the cross-correlation function, obtain the time offset corresponding to the maximum peak point, and determine it as the thermal response hysteresis time.

[0051] It should be noted that the initial measured temperature rise sequence was generated using vectorized point-to-point difference operations. The system reads the top oil temperature sequence. With the ambient temperature sequence For any point in time Calculate the difference This operation physically decouples the influence of ambient base temperature on the equipment's temperature rise, extracting the heat effect component purely caused by internal transformer losses.

[0052] It should be noted that the calculation of the cross-correlation function and the determination of the thermal response hysteresis time are achieved using the Normalized Cross-Correlation (NCC) algorithm. First, the load current sequence is defined as... The initial measured temperature rise sequence is as follows: Then, within the preset lag search interval... Within, for each trial time offset Calculate the cross-correlation coefficient using the following formula. :

[0053]

[0054] in, For sequence length, and These represent the mean values ​​of the data within the corresponding window. This formula measures the load current sequence during translation. After each time unit, the similarity to the temperature rise sequence waveform is calculated. Finally, all waveforms within the search interval are iterated over. Value, find the Reaching the global maximum value ,Right now:

[0055]

[0056] Should This refers to the determined thermal response hysteresis time, which physically means the average delay time for load changes to be transmitted to changes in the top layer oil temperature.

[0057] It is worth noting that the preset upper limit of the lag search interval... The determination was based on statistical analysis of historical temperature rise test data for transformers of the same model. The system collected a large number of thermal response curves of transformers under step loads, extracted the time constant distribution from the sudden load change to temperature rise stabilization, and selected the 99th percentile of this distribution (e.g., 180 minutes) as the baseline. This setting ensures that the search range covers the vast majority of possible physical lag scenarios, while avoiding computational redundancy and spurious peak interference caused by an excessively large search range.

[0058] For example, assume a sampling frequency of 1 minute per sampling. Load current sequence. exist A step increase occurs at a certain moment, while the initial measured temperature rise sequence... exist Only at a certain point does the system exhibit a corresponding significant upward trend. Cross-correlation calculations are performed within the specified range. At that time, the overlap between the shifted current sequence waveform and the temperature rise sequence waveform was the highest, and the cross-correlation coefficient was calculated. (Close to 1.0); while when At that time, the coefficient was only 0.2. Based on this, the system identified the offset corresponding to the peak value as 45, and determined the thermal response hysteresis time to be 45 minutes.

[0059] In step S13, the load current sequence is time-shifted according to the thermal response hysteresis time to determine the common effective time window of the shifted load current sequence, the ambient temperature sequence, and the initial measured temperature rise sequence. Synchronous data is then extracted based on the common effective time window to construct a time-aligned excitation input set and a time-aligned measured temperature rise sequence, including:

[0060] The load current sequence is shifted backward along the time axis by the thermal response hysteresis time to obtain the shifted load current sequence.

[0061] The overlapping time period between the effective start time point of the shifted load current sequence and the effective end time point of the initial measured temperature rise sequence is determined as the common effective time window;

[0062] Based on the start and end timestamps of the common effective time window, the shifted load current sequence and the ambient temperature sequence are extracted respectively and combined to obtain the time-aligned excitation input set;

[0063] Based on the start and end timestamps of the common effective time window, the initial measured temperature rise sequence is synchronously extracted to obtain the time-aligned measured temperature rise sequence.

[0064] It should be noted that shifting the load current sequence backward along the time axis by the thermal response hysteresis time is achieved using array index remapping technology. Assume the original load current sequence is... The thermal response hysteresis time is sampling points (i.e.) The backward shift operation refers to constructing a new sequence. , making This means that, The temperature rise at any moment is due to The current generated at any given moment. In physical storage, this manifests as adding the indices of the original current array as a whole. and fill the head One invalid value (NaN) or discard the end of the original sequence. Data points are used to maintain consistent length, thereby establishing a causal alignment of stimulus-response relationships on the time axis.

[0065] It should be noted that determining the common valid time window and performing synchronous data interception is based on the intersection operation in set theory. Let the valid time interval of the original collected data be... After translation, the effective data range of the load current sequence becomes (Assuming no fill value is provided) or (Depending on the alignment reference, here the alignment temperature rise is used, therefore the valid interval of the current data is shifted to the right). To ensure that all sequences involved in the calculation have the same length and corresponding physical times, the system calculates the intersection of the valid time intervals of each sequence:

[0066]

[0067] Typically, if the time axis of the temperature rise sequence is used as the reference (i.e., without shifting the temperature rise sequence), then the common effective time window is... The system uses slicing to synchronously extract the shifted load current sequence based on the start and end indices of the window. Original environmental temperature sequence and initial measured temperature rise sequence Ensure that the length of the three extracted sequences is the same. .

[0068] For example, assuming the original data length is... Points (24-hour data), sampling interval of 1 minute, lag time calculated in step S12 minutes (corresponding to) (Points). The original load current sequence index is... In the shifted load current sequence, the index... Invalid data (because the corresponding (No data was collected at any given time), valid data begins at index 45 (corresponding to the 0th point of the original current). The common valid time window is defined as the index interval. The system performs a capture operation, capturing the load current after the shift. Part (i.e., the original current) (partial), extracting ambient temperature Partially, the initial measured temperature rise is extracted. Partially. Finally, three aligned sequences, each with a length of 1395 points, are obtained. At this point, the... Among the data points, temperature rise Precisely corresponding to the current that generates it (i.e., the physical current 45 minutes ago) and the current ambient temperature This eliminated the time misalignment.

[0069] In step S14, the time-aligned excitation input set is substituted into a preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence, including:

[0070] Based on the load current data in the time-aligned excitation input set, the winding loss power and core heating power of the transformer are calculated, and a heat source input vector is constructed.

[0071] Obtain the thermal capacity coefficient matrix and thermal resistance network parameters of the transformer, and construct a set of first-order heat conduction differential equations;

[0072] The ambient temperature data in the excitation input set after aligning the heat source input vector with the time is substituted into the first-order heat conduction differential equation system for iterative solution to calculate the nodal temperature difference values ​​that change with time, thereby generating the theoretical reference temperature rise sequence.

[0073] It should be noted that the heat source input vector is constructed using a loss calculation formula based on load factor. The system first defines the rated loss parameters of the transformer, including the rated no-load loss. (Corresponding to core heating, considered constant) and rated load loss (Corresponding winding losses). For any moment in the time-aligned excitation input set. load current Calculate its actual total heating power :

[0074]

[0075] in This is the transformer's rated current. This constitutes the core component of the heat source input vector.

[0076] It should be noted that the construction and solution of the first-order heat conduction differential equations were achieved by discretizing the IEEE C57.91 standard thermal model using the forward Euler method. The preset transformer thermal circuit model follows the following energy balance differential equations:

[0077]

[0078] in, The top oil temperature, For ambient temperature, For equivalent heat capacity ( ), For equivalent thermal resistance ( To obtain the theoretical reference temperature rise sequence. The system transforms the above equations into a discrete iterative format. Let the sampling interval be... Then the first Theoretical temperature rise at time The calculation is as follows:

[0079]

[0080] The iterative process starts with the initial state of the sequence (usually set to 0 or a steady-state value) and gradually generates the complete theoretical reference temperature rise sequence.

[0081] It is worth noting that the heat capacity coefficient matrix... With thermal resistance network parameters The determination of the objective function is based on parameter identification of sample data from the transformer's historical healthy operation phases. Specifically, the system selects a period of peak data (e.g., 7 consecutive days of operating data) immediately after the transformer is put into operation or after maintenance to construct the objective function. :

[0082]

[0083] Using the least squares method to find the objective function Minimize the combination of parameters This data is then fixed as preset model parameters. This process is equivalent to training a unique digital twin thermal parameter model for each transformer, eliminating individual variation errors caused by universal parameters.

[0084] For example, suppose a large power transformer has a rated current Through the aforementioned parameter identification and training process, the device's unique thermal parameters, including its equivalent thermal resistance, are obtained. Equivalent heat capacity (This value conforms to the physical characteristics of large thermal inertia in large power transformers, ensuring a gradual temperature rise.) Set the sampling interval. Hour (1 minute). Rated load loss is known. Rated no-load loss In the current At what time, assuming an initial theoretical temperature rise At this point, the system acquires the time-aligned excitation input data and the load current. Calculate real-time heat generation power. According to the formula for the square of the load factor:

[0085]

[0086] Calculate real-time heat dissipation power According to the definition of thermal resistance, the heat dissipation power depends on the ratio of the current temperature rise to the thermal resistance:

[0087]

[0088] Calculate net calorie content The difference between the heating power and the heat dissipation power is the net power that causes the temperature rise.

[0089]

[0090] Iterative calculation of the next time step The theoretical temperature rise is determined using the forward Euler method discrete iterative formula:

[0091]

[0092] Calculate the temperature rise increment. ,but The calculation results show that, with a net heat power of 4kW, due to the extremely high heat capacity (200kWh / K) of the large transformer, its temperature rise within one minute is very small (approximately 0.00033K), which perfectly matches the smooth temperature rise evolution caused by the large thermal inertia during transformer operation. The system then iteratively calculates to generate a complete theoretical baseline temperature rise sequence, providing a high-precision physical benchmark for subsequent residual analysis.

[0093] In step S15, the time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence are subjected to time-domain difference operation to obtain a temperature rise residual sequence. The temperature rise residual sequence is then subjected to moving average filtering. A fault feature vector is constructed based on the processed temperature rise residual sequence, including:

[0094] The point-by-point difference between the time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence is calculated to obtain the original residual sequence;

[0095] The original residual sequence is smoothed and denoised using a moving average filter with a preset window width to obtain the processed temperature rise residual sequence.

[0096] Extract the mean deviation, fluctuation variance, peak density, and amplitude energy of the processed temperature rise residual sequence within the statistical period;

[0097] The fault feature vector is constructed by normalizing and combining the mean deviation, the fluctuation variance, the peak density, and the amplitude energy.

[0098] It should be noted that the processed temperature rise residual sequence is obtained by cascading time-domain differencing and linear smoothing filtering. First, let... The measured temperature rise after time alignment is The theoretical reference temperature rise is Calculate the original residuals:

[0099]

[0100] Next, construct a length of Sliding window (where (for single-sided span), Perform convolution smoothing:

[0101]

[0102] This operation generates This is the processed temperature rise residual sequence, which effectively filters out the high-frequency random white noise of the sensor itself and retains the low-frequency components that reflect the trend of thermal failure.

[0103] It should be noted that the construction of the aforementioned fault feature vector is based on multi-dimensional statistical moment analysis. For a length of... Residual fragments within the statistical period Calculate the mean bias of each of the following four physical characteristic components. Characterizes the overall deviation in temperature rise:

[0104]

[0105] Variance Characterizing the instability of temperature rise:

[0106]

[0107] Peak Density This characterizes the frequency of localized hotspots.

[0108]

[0109] in To satisfy local maxima and amplitude exceeding the preset noise floor The number of points; the noise floor can be obtained by analyzing the fluctuation standard deviation of the residual sequence under historical normal operating conditions, or set to an empirical small value (such as 0.1K) to filter out small fluctuations in the sensor.

[0110] Amplitude energy (Energy) Characterizes the total intensity of accumulated thermal faults:

[0111]

[0112] Finally, the Z-Score normalization method is used to map the four features to the same dimension, constructing a feature vector. :

[0113]

[0114] Subscript The parameters are the statistical mean and standard deviation of historical normal samples.

[0115] It is worth noting that the preset window width ( The determination of the frequency cutoff frequency (FFT) is based on spectral analysis of historical residual data. The system performs a Fast Fourier Transform (FFT) on the original residual sequence under normal operating conditions to identify the cutoff frequency of high-frequency noise. (e.g., 0.1Hz). Based on the sampling frequency. Select window width This ensures the filter has a -3dB cutoff characteristic. In this embodiment, the window width is set to 5 points (i.e., 5 minutes) based on statistical results.

[0116] For example, select a statistical period. Minutes. Calculations show that the characteristic value for a given time period is the mean deviation. ℃, fluctuation variance Peak density (i.e., 6 peaks), amplitude energy Historically, the normal reference value is: ; ; ; Substituting into the normalization formula, we obtain component 1 as... Component 2 is , component 3 is Component 4 is The final constructed fault feature vector This vector will be input into the subsequent classification model for diagnosis.

[0117] In step S16, the fault feature vector is input into a preset fault diagnosis classification model to determine the potential fault type label and fault severity level, including:

[0118] Input the fault feature vector into the preset fault diagnosis classification model, and calculate the geometric interval distance from the fault feature vector to the hyperplane of each fault category;

[0119] The classification confidence level is determined based on the geometric interval distance, and the category with the highest confidence level is identified as the potential fault type label.

[0120] The classification confidence score and the fault feature vector are normalized respectively, and the normalized classification confidence score and the normalized fault feature vector are weighted and summed to calculate the fault severity level.

[0121] It should be noted that the pre-set fault diagnosis classification model is constructed using a Support Vector Machine (SVM) based on a one-versus-rest (OvR) strategy. This model consists of multiple binary classifiers, each corresponding to a specific fault mode (such as winding overheating, multiple core grounding points, or oil passage blockage). For the input fault feature vector... , No. The decision function values ​​of the binary classifiers (i.e., the non-normalized form of the geometric margin distance). The calculation is as follows:

[0122]

[0123] in, For the set of support vectors, For Lagrange multipliers, For training sample labels, It is the radial basis function (RBF kernel). This is a bias term. This value... The absolute value of represents how far the current sample is from the classification hyperplane. The farther the distance, the greater the probability that it belongs to that category (or the probability that it does not).

[0124] It should be noted that the classification confidence level was determined using the Platt Scaling probability calibration method. The above decision function values... Mapped to posterior probability :

[0125]

[0126] in, and The Sigmoid parameters are obtained during the model training phase through maximum likelihood estimation. The system calculates the posterior probability of all failure modes and selects the category with the highest probability. As a label for the potential fault type:

[0127]

[0128] The corresponding maximum probability value at this time This refers to the classification confidence level.

[0129] It is worth noting that the specific training process and parameter settings of the fault diagnosis classification model are as follows: collect historical operating data of the transformer at different fault evolution stages, extract residual features to construct a fault feature vector library, and label the corresponding fault type. For the first... For each failure mode, samples belonging to that mode are denoted as positive, and all other modes and normal samples are denoted as negative. The optimal classification hyperplane is found by solving a quadratic programming problem. A radial basis function kernel is selected. Among them, kernel parameters With penalty factor Using 5-fold cross-validation A grid search is performed within the range to maximize the classification accuracy of the model on the validation set.

[0130] It should be noted that the quantified fault severity level is calculated using a multi-factor weighted fusion method. First, the classification confidence level... and amplitude energy features in fault feature vectors Perform Min-Max normalization; the amplitude energy feature Use the raw amplitude energy value calculated during the construction of the fault feature vector, rather than the normalized component in the feature vector:

[0131]

[0132] Since the probability itself is between 0 and 1, no additional processing is required, and it can be directly used as a normalized confidence index.

[0133]

[0134] in, and These are the minimum and maximum amplitude energy values ​​in the historical fault sample database, respectively. Then, the severity is calculated using a linear weighted formula. :

[0135]

[0136] in, and The preset weighting coefficients satisfy... The severity For one The dimensionless values ​​between these ranges indicate that the higher the value, the greater the certainty of the fault and the greater the destructive energy.

[0137] It is worth noting that the weighting coefficients and The determination was based on Pearson correlation analysis of hazard assessment data from historical failure cases. The system analyzed the correlation between failure confidence and the degree of temperature rise deviation caused by the failure in historical cases. And the correlation between fault energy and the degree of temperature rise deviation caused by the fault. Calculate the weights. , In this embodiment, the settings are based on statistical results. (Emphasis on diagnostic accuracy) (Focusing on fault intensity).

[0138] In one implementation, this embodiment presets a safety threshold (i.e., a yellow warning threshold). And the circuit breaker threshold (i.e., the red alert threshold). The security threshold The setting is based on Receiver Operating Characteristic (ROC) curve analysis of historical operation and maintenance data, selecting the severity value corresponding to the maximum Youden Index to ensure an optimal balance between sensitivity and specificity when identifying abnormal states. The circuit breaker threshold... The setting is based on the lower bound of the severity distribution of severe fault samples that must be immediately shut down for maintenance in historical data (such as the 95th percentile), aiming to minimize the risk of underreporting catastrophic accidents.

[0139] For example, using the data from the aforementioned steps, the fault feature vector Input the SVM model. Calculate the decision value for the winding overheating category. The confidence level is obtained through Platt Scaling mapping. For other categories, the confidence level was below 0.1. Therefore, the potential fault type was determined to be winding overheating, with a classification confidence level of [missing information]. The amplitude energy extracted in the aforementioned steps Maximum energy value in the historical fault database minimum value Normalized energy Calculate the severity level of the fault. If this value exceeds the preset yellow warning threshold (e.g., 0.6), the system will generate a warning instruction accordingly.

[0140] In step S17, if the severity level of the fault exceeds a preset safety threshold, a corresponding graded early warning instruction is generated based on the potential fault type label, including:

[0141] Map the severity level of the fault to a preset risk classification table;

[0142] If the mapping result falls within the low-risk range, a yellow alert is generated recommending enhanced monitoring.

[0143] If the mapping result falls within the high-risk range, a red warning instruction recommending shutdown and maintenance will be generated.

[0144] It should be noted that mapping the fault severity level to the preset risk classification table is achieved using an interval threshold determination algorithm. The system presets two key thresholds, a safety threshold, and a safety threshold. and circuit breaker threshold (satisfy For the fault severity level calculated in step S16 ,like If it is determined to be in a "normal state", no warning will be triggered. This is classified as "mild risk" and mapped to the yellow zone in the risk classification table. If... It was determined to be "severe risk" and mapped to the red zone in the risk classification table.

[0145] It should be noted that the generation of corresponding tiered warning instructions is achieved using rule-based template filling technology. The system retrieves a pre-set instruction template library, extracts the potential fault type label (such as "Type A") and severity level (numerical value), and fills it into the template corresponding to the risk level. For example, the yellow warning template is "[Fault Label] signs detected, severity [numerical value], it is recommended to execute [Strategy A]." The red warning template is "[Fault Label] serious risk detected, severity [numerical value], it is recommended to immediately execute [Strategy B]."

[0146] It is worth noting that the specific construction method of the pre-set instruction template library is as follows: the rule base is established based on the power transformer operation and maintenance regulations (such as DL / T 596), and a differentiated set of operation and maintenance suggestion strategies is preset for each potential fault type (such as winding overheating, core grounding, etc.). Each template adopts a four-segment configuration of "risk identifier + fault label + quantified severity + expert suggestion strategy". The system retrieves the corresponding handling suggestions from the strategy set according to the determined potential fault type label. For example, the mild risk strategy for "winding overheating" is "shorten the sampling cycle and strengthen oil chromatography analysis", and the severe risk strategy is "immediately apply for dispatch shutdown and conduct cover inspection".

[0147] It is worth noting that the security threshold With circuit breaker threshold The determination of severity is based on the analysis of Receiver Operating Characteristic (ROC) curves from historical maintenance data. The system collects historical samples and their final maintenance status (no faults, minor faults requiring monitoring, severe faults requiring shutdown), categorized by severity. Plot an ROC curve for the variable. Select the point corresponding to the maximum Youden Index. Value as This ensures a balance between sensitivity and specificity in detecting abnormal states. The point corresponding to the point with the minimum weighted overlap between the false positive and false negative probability distributions is selected. Value as This typically corresponds to the lower limit of the severity of samples in historical data that require immediate downtime for maintenance (e.g., the 95th percentile).

[0148] For example, continuing from the previous embodiments, the calculated fault severity level The potential fault type is labeled "winding overheat". Preset safety threshold. Circuit breaker threshold The system performs interval determination because... The mapping result is determined to be in the mild risk range. The system invokes the yellow warning logic to generate the instruction: "Yellow warning: Overheating risk detected in transformer windings, current severity level 0.752. It is recommended to shorten the sampling period to 30 seconds and enhance oil chromatography analysis." If the calculated... for If the risk is high, a red alert is generated: "A serious risk of overheating in the transformer windings has been detected. The current severity level is 0.85. It is recommended to immediately request a dispatch shutdown and conduct a cover inspection."

[0149] It should be noted that the generated graded early warning instructions (such as yellow or red warnings) will be sent to the transformer's monitoring host computer system or operation and maintenance management platform. This system or platform can perform one or more of the following operations, which will be highlighted in the corresponding color and displayed in a pop-up window on the graphical interface: storing the warning information, fault type, severity, and timestamp into the database and generating a log; sending alarm notifications to designated operation and maintenance personnel through preset communication interfaces (such as SMS gateway, WeChat / DingTalk robot, email server); and in the case of a red warning, further sending a signal of equipment status deterioration to the power grid dispatching system to provide a reference for dispatching decisions.

[0150] In summary, this invention achieves strict time-domain alignment between load excitation and temperature rise response by acquiring real-time multidimensional state sequences of transformers and accurately quantifying thermal response lag time using cross-correlation analysis. This effectively overcomes the problem of asynchrony between dynamic and static states caused by large thermal inertia in traditional monitoring. By combining a physical model based on thermal balance differential equations with moving average filtering technology, it achieves dual decoupling of the ambient temperature base and random noise, extracting high-confidence temperature rise residual features from dynamically changing measured data. Furthermore, it utilizes a fault diagnosis classification model to intelligently analyze and assess the severity of feature vectors, achieving accurate qualitative and quantitative fault classification. This invention thus constructs a complete closed-loop system from data cleaning to decision-making and early warning, significantly improving the real-time performance, accuracy, and intelligence level of fault prediction for power transformers under dynamic and complex operating conditions.

[0151] Reference Figure 2 The second embodiment of the present invention provides a power transformer fault early warning system, comprising:

[0152] The data acquisition module is used to acquire real-time operating status data of the transformer, including load current sequence, ambient temperature sequence and top oil temperature sequence.

[0153] The time series analysis module is used to perform difference calculations based on the top oil temperature sequence and the ambient temperature sequence to obtain the initial measured temperature rise sequence, and to perform cross-correlation analysis on the load current sequence and the initial measured temperature rise sequence to determine the thermal response hysteresis time.

[0154] The data alignment module is used to perform time-series shifting of the load current sequence based on the thermal response hysteresis time, determine the common effective time window of the shifted load current sequence, the ambient temperature sequence, and the initial measured temperature rise sequence, and perform synchronous data interception based on the common effective time window to construct the time-aligned excitation input set and the time-aligned measured temperature rise sequence.

[0155] The model calculation module is used to substitute the time-aligned excitation input set into a preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence.

[0156] The feature extraction module is used to perform time-domain difference operation on the time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence to obtain a temperature rise residual sequence, and to perform moving average filtering on the temperature rise residual sequence to construct a fault feature vector based on the processed temperature rise residual sequence.

[0157] The diagnostic decision module is used to input the fault feature vector into a preset fault diagnosis classification model to determine the potential fault type label and fault severity level;

[0158] The early warning response module is used to generate a corresponding graded early warning instruction based on the potential fault type label if the severity level of the fault exceeds a preset safety threshold.

[0159] It should be noted that the power transformer fault early warning system provided in this embodiment of the invention is used to execute all the process steps of the power transformer fault early warning method in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.

[0160] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a power transformer fault early warning program. When the processor executes the computer program, it implements the steps described in the various embodiments of the power transformer fault early warning method, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the data acquisition module.

[0161] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0162] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0163] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0164] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0165] If the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0166] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0167] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that 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 for those skilled in the art.

Claims

1. A method for early warning of power transformer faults, characterized in that, include: Collect real-time operating status data of the transformer, including load current sequence, ambient temperature sequence and top oil temperature sequence; The initial measured temperature rise sequence is obtained by performing a difference calculation between the top oil temperature sequence and the ambient temperature sequence, and a cross-correlation analysis is performed between the load current sequence and the initial measured temperature rise sequence to determine the thermal response hysteresis time. The load current sequence is time-shifted according to the thermal response hysteresis time to determine the common effective time window of the shifted load current sequence, the ambient temperature sequence, and the initial measured temperature rise sequence. Synchronous data is extracted based on the common effective time window to construct the time-aligned excitation input set and the time-aligned measured temperature rise sequence. The time-aligned excitation input set is substituted into a preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence. The time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence are subjected to time-domain difference operation to obtain the temperature rise residual sequence. The temperature rise residual sequence is then subjected to moving average filtering. A fault feature vector is constructed based on the processed temperature rise residual sequence. The fault feature vector is input into a preset fault diagnosis classification model to determine the potential fault type label and fault severity level; If the severity level of the fault exceeds a preset safety threshold, a corresponding graded early warning instruction is generated based on the potential fault type label. The step of substituting the time-aligned excitation input set into a preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence includes: Based on the load current data in the time-aligned excitation input set, the winding loss power and core heating power of the transformer are calculated, and a heat source input vector is constructed. Obtain the thermal capacity coefficient matrix and thermal resistance network parameters of the transformer, and construct a set of first-order heat conduction differential equations; The ambient temperature data in the excitation input set after aligning the heat source input vector with the time is substituted into the first-order heat conduction differential equation system for iterative solution to calculate the nodal temperature difference values ​​that change with time, thereby generating the theoretical reference temperature rise sequence.

2. The power transformer fault early warning method according to claim 1, characterized in that, The real-time operating status data of the acquired transformer includes load current sequence, ambient temperature sequence, and top oil temperature sequence, including: The transformer's top oil temperature, ambient temperature, and load current are continuously and synchronously collected using temperature sensors and current transformers at a preset sampling frequency. The collected continuous numerical points are arranged in order of collection timestamp to generate a raw data stream containing the time dimension; The original data stream is preprocessed by imputing missing values ​​and removing outliers, and the load current sequence, the ambient temperature sequence, and the top oil temperature sequence are extracted by column.

3. The power transformer fault early warning method according to claim 1, characterized in that, The process involves performing a difference calculation between the top-layer oil temperature sequence and the ambient temperature sequence to obtain an initial measured temperature rise sequence, and then performing a cross-correlation analysis between the load current sequence and the initial measured temperature rise sequence to determine the thermal response hysteresis time, including: Subtract the corresponding value in the ambient temperature sequence from the value at each moment in the top oil temperature sequence to generate the initial measured temperature rise sequence; Set a time sliding window and calculate the cross-correlation function between the load current sequence and the initial measured temperature rise sequence at different time offsets; Search for the maximum peak point of the cross-correlation function, obtain the time offset corresponding to the maximum peak point, and determine it as the thermal response hysteresis time.

4. The power transformer fault early warning method according to claim 1, characterized in that, The step of time-shifting the load current sequence based on the thermal response hysteresis time, determining the common effective time window of the shifted load current sequence, the ambient temperature sequence, and the initial measured temperature rise sequence, and performing synchronous data extraction based on the common effective time window to construct a time-aligned excitation input set and a time-aligned measured temperature rise sequence includes: The load current sequence is shifted backward along the time axis by the thermal response hysteresis time to obtain the shifted load current sequence. The overlapping time period between the effective start time point of the shifted load current sequence and the effective end time point of the initial measured temperature rise sequence is determined as the common effective time window; Based on the start and end timestamps of the common effective time window, the shifted load current sequence and the ambient temperature sequence are extracted respectively and combined to obtain the time-aligned excitation input set; Based on the start and end timestamps of the common effective time window, the initial measured temperature rise sequence is synchronously extracted to obtain the time-aligned measured temperature rise sequence.

5. The power transformer fault early warning method according to claim 1, characterized in that, The process involves performing a time-domain difference operation between the time-aligned measured temperature rise sequence and the theoretical baseline temperature rise sequence to obtain a temperature rise residual sequence. This residual sequence is then subjected to a moving average filter. A fault feature vector is constructed based on the processed temperature rise residual sequence, including: The point-by-point difference between the time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence is calculated to obtain the original residual sequence; The original residual sequence is smoothed and denoised using a moving average filter with a preset window width to obtain the processed temperature rise residual sequence. Extract the mean deviation, fluctuation variance, peak density, and amplitude energy of the processed temperature rise residual sequence within the statistical period; The fault feature vector is constructed by normalizing and combining the mean deviation, the fluctuation variance, the peak density, and the amplitude energy.

6. The power transformer fault early warning method according to claim 1, characterized in that, The step of inputting the fault feature vector into a preset fault diagnosis classification model to determine the potential fault type label and fault severity level includes: Input the fault feature vector into the preset fault diagnosis classification model, and calculate the geometric interval distance from the fault feature vector to the hyperplane of each fault category; The classification confidence level is determined based on the geometric interval distance, and the category with the highest confidence level is identified as the potential fault type label. The classification confidence score and the fault feature vector are normalized respectively, and the normalized classification confidence score and the normalized fault feature vector are weighted and summed to calculate the fault severity level.

7. The power transformer fault early warning method according to claim 1, characterized in that, If the severity level of the fault exceeds a preset safety threshold, a corresponding graded early warning instruction is generated based on the potential fault type label, including: Map the severity level of the fault to a preset risk classification table; If the mapping result falls within the low-risk range, a yellow alert is generated recommending enhanced monitoring. If the mapping result falls within the high-risk range, a red warning instruction recommending shutdown and maintenance will be generated.

8. A power transformer fault early warning system, characterized in that, include: The data acquisition module is used to acquire real-time operating status data of the transformer, including load current sequence, ambient temperature sequence and top oil temperature sequence. The time series analysis module is used to perform difference calculations based on the top oil temperature sequence and the ambient temperature sequence to obtain the initial measured temperature rise sequence, and to perform cross-correlation analysis on the load current sequence and the initial measured temperature rise sequence to determine the thermal response hysteresis time. The data alignment module is used to perform time-series shifting of the load current sequence based on the thermal response hysteresis time, determine the common effective time window of the shifted load current sequence, the ambient temperature sequence, and the initial measured temperature rise sequence, and perform synchronous data interception based on the common effective time window to construct the time-aligned excitation input set and the time-aligned measured temperature rise sequence. The model calculation module is used to substitute the time-aligned excitation input set into a preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence. The feature extraction module is used to perform time-domain difference operation on the time-aligned measured temperature rise sequence and the theoretical reference temperature rise sequence to obtain a temperature rise residual sequence, and to perform moving average filtering on the temperature rise residual sequence to construct a fault feature vector based on the processed temperature rise residual sequence. The diagnostic decision module is used to input the fault feature vector into a preset fault diagnosis classification model to determine the potential fault type label and fault severity level; The early warning response module is used to generate a corresponding graded early warning instruction based on the potential fault type label if the severity level of the fault exceeds a preset safety threshold. The step of substituting the time-aligned excitation input set into a preset transformer thermal circuit model based on thermal balance differential equations to calculate the theoretical reference temperature rise sequence includes: Based on the load current data in the time-aligned excitation input set, the winding loss power and core heating power of the transformer are calculated, and a heat source input vector is constructed. Obtain the thermal capacity coefficient matrix and thermal resistance network parameters of the transformer, and construct a set of first-order heat conduction differential equations; The ambient temperature data in the excitation input set after aligning the heat source input vector with the time is substituted into the first-order heat conduction differential equation system for iterative solution to calculate the nodal temperature difference values ​​that change with time, thereby generating the theoretical reference temperature rise sequence.

Images