A power transformer full life cycle cost prediction method and system

Abstract

The application discloses a kind of electric power transformer full life cycle cost prediction method and system, method includes: collection oil dissolved gas historical concentration and cost parameter;Gas sequence is decomposed and reconstructed to CEEMDAN;Input LSTM model predicts future gas concentration;Splice history and predicted sequence and normalize;Calculate no coding ratio and carry out fault diagnosis;State transition analysis is carried out to fault type time sequence, adjacent similar fault section is merged, and similar fault section with interval not more than recovery threshold is bridged and merged to obtain fault event sequence;Fault type is mapped to A, B type fault and the number of days, event frequency and maximum cumulative duration are counted;Dynamic fault cost is calculated;With initial total cost is summarized and output full life cycle cost.The application can dynamically and accurately predict full life cycle cost, establish complete mapping from gas concentration to fault event to cost, improve prediction accuracy and engineering practicability.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition assessment and intelligent operation and maintenance technology, specifically to a method and system for predicting the full life cycle cost of power transformers. Background Technology

[0002] As a key piece of equipment in the power system, the operating status of power transformers directly affects the reliability of power supply and the economics of operation and maintenance. Predicting and assessing the entire life-cycle cost of power transformers is beneficial for enterprises to optimize their overall planning in equipment procurement, operation and maintenance, overhaul decisions, and decommissioning arrangements.

[0003] Most existing lifecycle cost assessment schemes focus on relatively fixed cost items such as initial investment, operation, maintenance, and decommissioning. However, they typically estimate failure costs using empirical values, static parameters, or post-hoc statistical results, which fails to fully reflect the impact of future changes in equipment condition on failure costs. This static estimation method cannot capture the potential failure evolution trends that may occur during equipment operation, leading to significant discrepancies between cost predictions and actual operating conditions.

[0004] Meanwhile, transformer fault prediction and diagnosis technologies based on dissolved gas analysis in transformer oil have been extensively studied. However, existing solutions often remain at the level of gas concentration prediction, single-moment fault identification, or independent cost assessment, lacking an integrated technical process that constructs a fault event sequence along the time axis from future multi-gas sequence diagnostic results and further drives fault cost calculation. Specifically, existing technologies typically treat gas concentration prediction, fault diagnosis, and cost assessment as independent steps, failing to establish a complete mapping relationship from gas concentration changes to fault event evolution and cost impact. Regarding gas concentration prediction, although various time-series prediction models have been applied to predict dissolved gas concentrations in transformer oil, these prediction results are often only used for single-moment fault diagnosis, failing to integrate multi-gas prediction sequences into a continuous fault evolution trend, and thus unable to provide dynamic input for full life-cycle cost prediction.

[0005] Furthermore, when directly statistically analyzing fault results by time unit, short-term fluctuations can easily fragment similar faults into multiple segments, thus affecting the stability of fault frequency, duration, and cost quantification results. For example, when a transformer experiences a high-temperature overheating fault and briefly recovers, only to subsequently experience a similar fault, existing technologies often treat these two faults as independent events, when in reality they may belong to different stages of the same fault process. This fragmented fault statistics method leads to an overestimation of fault event frequency and an underestimation of fault duration, thereby affecting the accuracy of fault cost calculation. Existing technologies lack a bridging and merging mechanism based on recovery thresholds, failing to identify similar faults before and after short-term recovery as the same fault event, resulting in fault event statistics that do not match the actual operating state of the equipment.

[0006] More importantly, existing technologies fail to establish an effective correlation between fault type, fault event characteristics, and fault cost. While dissolved gas analysis in oil can identify various fault types, such as low-temperature overheating, medium-temperature overheating, high-temperature overheating, partial discharge, spark discharge, and arc discharge, these fault types exhibit significant differences in the degree of damage to equipment and repair costs. Existing cost assessment methods typically treat all fault types as equally severe or simply classify them based on fault type, failing to consider the impact of dynamic characteristics such as the frequency and duration of fault events on costs. This results in a lack of accuracy and foresight in fault cost estimation.

[0007] Therefore, a technical solution is needed that can unify the future fault evolution trend of transformers, fault type time series results, fault event construction, severity mapping and full life cycle economic calculation, to realize a complete technical chain from multi-gas sequence prediction to fault cost quantification, and provide more accurate and reliable technical support for the full life cycle cost prediction of power transformers. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for predicting the full life cycle cost of power transformers. This method and system can achieve dynamic and accurate prediction of the full life cycle cost of power transformers, establish a complete mapping relationship from changes in dissolved gas concentration in oil to the evolution of fault events and the quantification of fault costs, effectively improve the consistency between cost prediction results and actual operating conditions, and provide a reliable basis for power transformer operation and maintenance decisions.

[0009] In a first aspect, the present invention provides a method for predicting the full life-cycle cost of a power transformer, comprising the following steps: S1. Collect multi-source data of the power transformer during operation; the multi-source data includes: historical gas concentration sequence data of various target gases dissolved in oil and full life cycle cost parameters; S2. Preprocess the historical gas concentration sequence data of each target gas, and perform CEEMDAN decomposition and reconstruction on the preprocessed historical gas concentration sequence data to obtain the reconstructed input sequence of each target gas. S3. Input the reconstructed input sequences of each target gas into the LSTM model to obtain the predicted gas concentration sequence within the future target prediction period. S4. The historical gas concentration sequence and the predicted gas concentration sequence are spliced ​​together according to a unified time axis to form a fused gas concentration sequence covering the target life cycle, and the fused gas concentration sequence is subjected to Min-Max normalization. S5. Calculate the uncoded ratio based on the fused gas concentration sequence after Min-Max normalization, and perform fault diagnosis according to the preset time unit to obtain the fault type time series results of the power transformer within the target life cycle. S6. Perform state transition analysis on the time sequence results of the fault type, merge the time units of adjacent similar faults into continuous fault segments, and perform bridging and merging on the time units of similar faults with an interval not exceeding a preset recovery threshold to obtain a fault event sequence. S7. Map the fault types in the fault event sequence to type A faults and type B faults, and count the number of days, frequency of fault events and maximum cumulative duration of type A and type B faults in the fault event sequence respectively. S8. Calculate the dynamic failure cost based on the statistical results of S7; S9. Summarize the dynamic fault cost with the initial total cost to output the full life cycle cost of the power transformer.

[0010] Secondly, the present invention provides a power transformer full life cycle cost prediction system for implementing the above method, comprising: The data acquisition module is used to collect multi-source data of the power transformer during operation; the multi-source data includes: historical gas concentration sequence data of various target gases dissolved in oil and full life cycle cost parameters; The data preprocessing module is used to preprocess the historical gas concentration sequence data of each target gas, and to perform CEEMDAN decomposition and reconstruction on the preprocessed historical gas concentration sequence data to obtain the reconstructed input sequence of each target gas. The decomposition prediction module is used to input the reconstructed input sequences of each target gas into the LSTM model to obtain the predicted gas concentration sequence within the future target prediction period; The fusion sequence construction module is used to splice historical gas concentration sequences and predicted gas concentration sequences according to a unified time axis to form a fusion gas concentration sequence covering the target life cycle, and to perform Min-Max normalization processing on the fusion gas concentration sequence. The fault diagnosis module is used to calculate the uncoded ratio based on the fused gas concentration sequence after Min-Max normalization, and to perform fault diagnosis according to the preset time unit to obtain the fault type time series results of the power transformer within the target life cycle. The state transition analysis and event construction module is used to perform state transition analysis on the time series results of the fault type, merge the time units of adjacent similar faults into continuous fault segments, and perform bridging and merging on the time units of similar faults with an interval not exceeding a preset recovery threshold to obtain a fault event sequence. The severity mapping module is used to map the fault types in the fault event sequence to Class A faults and Class B faults, and to count the number of days, frequency of fault events and maximum cumulative duration of Class A and Class B faults in the fault event sequence, respectively. The failure cost calculation module is used to calculate dynamic failure costs based on the statistical results output by the severity mapping module. The full life cycle cost output module is used to summarize the dynamic fault cost and the initial total cost, and output the full life cycle cost of the power transformer.

[0011] As can be seen from the above, the power transformer full life cycle cost prediction method and system provided in this application collects multi-source operating data and sequentially completes gas sequence preprocessing decomposition and reconstruction, future gas concentration prediction, fusion sequence construction, time-series fault diagnosis, fault event merging construction, and fault statistics. Finally, it calculates dynamic fault costs and summarizes them to obtain the full life cycle cost. It can establish a complete mapping relationship from changes in dissolved gas concentration in oil to fault event evolution and then to fault cost quantification. It solves the problems of large deviations in static estimation of fault costs and inaccurate cost calculation caused by fragmented fault statistics in the existing technology. Attached Figure Description

[0012] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a method for predicting the full life-cycle cost of a power transformer according to an embodiment of the present invention. Figure 2 This is a logical framework diagram of a method for predicting the full life cycle cost of a power transformer according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a power transformer life cycle cost prediction system according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the lifecycle cost composition, including dynamic failure costs, according to an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0015] Traditional life-cycle cost assessment schemes for power transformers typically estimate failure costs using empirical values, static parameters, or post-hoc statistical results, failing to fully reflect the dynamic impact of future equipment condition changes on failure costs. Furthermore, existing failure prediction and diagnosis technologies based on dissolved gas analysis lack an integrated process for constructing a failure event sequence from future multi-gas sequence diagnostic results along a time axis to further drive failure cost calculation. In addition, directly statistically analyzing failure results by time unit can lead to short-term fluctuations fragmenting similar failures into multiple isolated segments, affecting the stability of failure frequency, duration, and cost quantification results.

[0016] In this regard, such as Figures 1-2 As shown, this application proposes a method for predicting the full life-cycle cost of power transformers, including the following steps: Step S1: Collect multi-source data of the power transformer during operation; the multi-source data includes: historical gas concentration sequence data of various target gases dissolved in oil and full life cycle cost parameters; Step S2: Preprocess the historical gas concentration sequence data of each target gas, and perform CEEMDAN decomposition and reconstruction on the preprocessed historical gas concentration sequence data to obtain the reconstructed input sequence of each target gas. Step S3: Input the reconstructed input sequences of each target gas into the LSTM model to obtain the predicted gas concentration sequence within the future target prediction period; Step S4: The historical gas concentration sequence and the predicted gas concentration sequence are spliced ​​together according to a unified time axis to form a fused gas concentration sequence covering the target lifetime, and the fused gas concentration sequence is subjected to Min-Max normalization. Step S5: Calculate the uncoded ratio based on the fused gas concentration sequence after Min-Max normalization, and perform fault diagnosis according to the preset time unit to obtain the fault type time series results of the power transformer within the target life cycle. Step S6: Perform state transition analysis on the fault type time series results, merge the time units of adjacent similar faults into continuous fault segments, and perform bridging and merging on the time units of similar faults with an interval not exceeding the preset recovery threshold to obtain the fault event sequence. Step S7: Map the fault types in the fault event sequence to type A faults and type B faults, and count the number of days, frequency of fault events and maximum cumulative duration of type A and type B faults in the fault event sequence respectively. Step S8: Calculate the dynamic failure cost based on the statistical results of step S7; Step S9: Summarize the dynamic fault cost with the initial total cost to output the full life cycle cost of the power transformer.

[0017] Multi-source data refers to various types of data related to the operating status and cost of power transformers obtained from different sources, such as sensor monitoring data, equipment file data, and financial data.

[0018] Historical gas concentration sequence data refers to a data set formed by arranging the concentration values ​​of various target gases (such as H2, CH4, etc.) dissolved in the oil in the power transformer in chronological order at different time points during the operation of the power transformer.

[0019] The life cycle cost parameter refers to the parameters used to calculate all relevant costs of a power transformer throughout its entire life cycle, from procurement to decommissioning, including but not limited to investment, operation, maintenance, decommissioning, and fault repair costs.

[0020] CEEMDAN decomposition and reconstruction is a signal processing method used to decompose complex nonlinear nonstationary time series into multiple intrinsic mode function components with different frequency characteristics and a residual component. By selectively superimposing these components, a more stationary and regular sequence can be reconstructed to remove noise and highlight the main trend.

[0021] LSTM models are a special type of recurrent neural network that excels at processing and predicting time series data. Their internal structure enables them to learn long-term dependencies, effectively capturing the temporal dynamics of sequence data.

[0022] Min-Max normalization is a data preprocessing technique that maps raw data to a specific interval through linear transformation, eliminating the influence of dimensions and facilitating comparison and model processing of data with different dimensions.

[0023] Uncoded ratios refer to a method of diagnosing transformer fault types by calculating specific ratios between the concentrations of dissolved gases in the oil. These ratios are typically associated with specific fault modes and do not require complex coding rules.

[0024] Fault type timing results refer to the sequence of fault types arranged in chronological order, obtained through fault diagnosis of each time unit within the target life cycle of a power transformer.

[0025] State transition analysis refers to the analysis of the timing results of fault types to identify the changing patterns of fault states, such as from a normal state to a fault state, or from one fault type to another.

[0026] Bridging and merging refers to the operation in state transition analysis where two faulty segments of the same type are treated as the same fault event and connected and merged when there is a brief normal interval between them.

[0027] A fault event sequence refers to a set of independent fault events arranged in chronological order, which are integrated by state transition analysis and bridging merging of consecutive or bridged fault time units of the same type.

[0028] Class A and Class B faults refer to different types of faults classified according to their severity or impact on equipment operation. For example, Class A faults may represent relatively minor or controllable faults, while Class B faults may represent more serious or urgent faults.

[0029] Dynamic failure cost refers to the failure-related cost that varies with the equipment status, calculated based on statistics such as the actual failure event sequence, failure type, number of days, event frequency, and cumulative duration of power transformers within the target life cycle, combined with corresponding maintenance costs.

[0030] Initial total cost refers to all costs incurred by a power transformer before it is put into operation, typically including investment costs, operating costs, maintenance costs, and decommissioning costs.

[0031] The total life cycle cost refers to the total cost incurred by a power transformer throughout its entire life cycle, from planning, design, procurement, installation, operation, maintenance to final decommissioning and disposal. It is the sum of the initial total cost and the dynamic failure cost.

[0032] This application overcomes the static and lagging problems of fault cost estimation in traditional methods by fusing and diagnosing the historical and predicted sequences of dissolved gases in power transformer oil, and further converting the time-series results of fault types into fault event sequences. Through severity mapping and dynamic cost calculation of fault events, this method can provide forward-looking predictions of the entire life cycle cost, thus providing a basis for decision-making in power transformer maintenance strategy formulation, budget allocation, and asset management.

[0033] In one optional implementation, the target gas includes H2, CH4, C2H2, C2H4, C2H6, CO, and CO2; the life cycle cost parameters include investment cost, operating cost, maintenance cost, decommissioning cost, average repair cost for Class A faults, and average repair cost for Class B faults; wherein, investment cost includes equipment purchase cost and installation cost; operating cost includes operating loss cost and switching cost.

[0034] Specifically, the selected H2, CH4, C2H2, C2H4, C2H6, CO, and CO2 are recognized key fault characteristic gases in dissolved gas analysis (DGA) of power transformer oil. By comprehensively collecting the concentration data of these core gases, sufficient and accurate basis can be provided for subsequent fault diagnosis, ensuring the reliability of fault type determination.

[0035] Meanwhile, the defined life-cycle cost parameters constitute the complete cost chain of power transformers from purchase, operation, maintenance to decommissioning. Investment cost refers to the initial fixed expenditure before the transformer is put into use, specifically including equipment purchase cost and installation cost. Equipment purchase cost refers to the cost of purchasing the power transformer itself and its supporting equipment (such as cooling systems, bushings, tap changers, etc.). Installation cost refers to the labor, material, and machinery costs required to transport the transformer to the site and perform assembly, commissioning, and grid connection. Operating cost refers to the continuous expenditure during the transformer's service life, specifically including operating loss cost and switching cost. Operating loss cost refers to the cost incurred due to power loss caused by iron and copper losses during normal operation. Switching cost refers to the power generation loss, grid dispatching costs, and indirect economic losses affecting the reliability of power supply to users caused by power outages due to transformer failure or maintenance. Maintenance cost refers to the costs incurred during transformer operation for preventative maintenance, planned overhauls, and component replacement to maintain its normal operating condition and extend its service life, including labor costs, material costs, and testing fees. Decommissioning costs refer to the expenses incurred when a transformer reaches its design life or is decommissioned prematurely, including dismantling, transportation, waste disposal (such as waste oil, scrap metal, and waste insulation materials), and site restoration. The average repair cost for Class A and Class B faults is set based on historical experience or industry standards to quantify the average repair expenditure after a fault of different severity (Class A faults typically refer to minor or repairable faults, while Class B faults typically refer to more serious faults or those leading to outages). These costs are key components in calculating dynamic fault costs. By comprehensively considering these cost parameters, a complete and refined cost prediction model can be constructed, avoiding the omission of key cost items and thus improving the accuracy and practicality of life-cycle cost prediction.

[0036] In one optional implementation, the historical gas concentration sequence data of each target gas is preprocessed, including missing value completion, outlier removal, and time alignment.

[0037] Specifically, missing value imputation aims to address the problem of missing data points in a data sequence due to various reasons (such as sensor failure, data transmission interruption, etc.). In practical applications, various methods can be used for missing value imputation. For example, for time series data, linear interpolation, spline interpolation, or prediction models based on historical trends can be used to imput missing values ​​to ensure the continuity and integrity of the data sequence. Furthermore, based on the statistical characteristics of the data, the mean, median, or mode can be used for imputation, or machine learning algorithms (such as K-nearest neighbors, matrix factorization, etc.) can be used to estimate missing values, thereby preserving the effective information of the original data to the greatest extent possible.

[0038] Outlier removal is used to identify, remove, or correct outlier data points in a data sequence that significantly deviate from the normal range. These outliers may originate from momentary sensor malfunctions, environmental interference, or data recording errors. If left untreated, they will negatively impact subsequent data analysis and model training. Outlier removal methods include, but are not limited to, statistical methods such as the 3σ criterion, interquartile range (IQR), or Z-score, which use thresholds to determine whether a data point is an outlier. Model-based methods can also be used; for example, training a predictive model to identify data points that deviate significantly from the model's predictions as outliers, or using unsupervised learning algorithms such as Isolation Forest to detect outliers in the data. When removing outliers, they can be directly deleted or replaced with adjacent normal values, the mean of the sequence, or the median to maintain the length and structure of the data sequence.

[0039] Time alignment aims to address the inconsistency in timestamps between gas concentration data from different sources or of different types. Due to potential differences in sampling frequencies, clock drift, or asynchronous recording in data acquisition systems, the time points of various gas concentration sequences cannot be precisely matched. Time alignment unifies all gas concentration data onto a common time axis. This is typically achieved through resampling techniques, such as resampling all data to a least common multiple of the time granularity, or selecting a standard time interval (e.g., hourly, daily) for unification. During resampling, missing data at new time points can be handled using interpolation (e.g., linear interpolation, nearest neighbor interpolation) or forward / backward padding methods to ensure that all gas concentration sequences are synchronized and comparable in the time dimension.

[0040] In one optional implementation, a specific method for CEEMDAN decomposition and reconstruction of the preprocessed historical gas concentration sequence data of each target gas is described. Specifically, firstly, the preprocessed historical gas concentration sequence data of each target gas is decomposed and reconstructed... X , where X represents , For the first tThe target gas concentration values ​​corresponding to each time point t The value range is 1 to T , T Given the total number of historical time points, perform CEEMDAN decomposition on each point. This decomposition process can transform the original complex, nonlinear, and non-stationary time series data. X Adaptively decomposed into N Vectors of intrinsic mode function components arranged in descending order of frequency. and a residual component vector Its expression is: ; in, Indicates the first i Each intrinsic mode function component is specifically: In order to be in t Historical concentration sequence data at each time point X The middle belongs to the first i The amplitude of each intrinsic mode function component; R Represents the residual component, specifically , exist t At each point in time, representing historical concentration sequence data X The sum of the remaining components after decomposition; N This represents the total number of intrinsic mode function components obtained after decomposition. CEEMDAN decomposition effectively solves the mode aliasing problem in traditional empirical mode decomposition (EMD) by introducing adaptive noise, resulting in each intrinsic mode function component having better physical meaning and independence.

[0041] Based on this, in order to further remove noise from the data, this application identifies and removes the preceding data that characterizes high-frequency random noise. k There are 10 intrinsic mode function components, of which k The value range is 1 to N -1. Typically, the most frequent intrinsic mode function components often contain a large amount of random noise or transient fluctuations. These components contribute little to long-term trend prediction and may even introduce interference. By setting appropriate... k The value can effectively separate these high-frequency noise components from the original signal. k The value can be determined based on methods such as empirical judgment, energy analysis, or correlation analysis.

[0042] Then, the remaining N - k The low-to-mid-frequency intrinsic mode function components are superimposed with the residual components to obtain the reconstructed input sequences of each target gas. Among them, the first tReconstruction values ​​at each time point The expression is: .

[0043] The reconstructed sequence retains the main trends and periodic characteristics of the original signal while significantly reducing high-frequency noise interference, making the data smoother and more stable, and providing high-quality input for subsequent LSTM model prediction.

[0044] In one alternative implementation, the Min-Max normalization process is expressed as: in, The original fusion gas concentration sequence data to be normalized. and These are the minimum and maximum values ​​in the original fusion gas concentration sequence, respectively. To prevent the use of a preset minimum positive number with a denominator of 0.

[0045] Specifically, Min-Max normalization is a commonly used data preprocessing technique that aims to linearly scale the raw data to a preset range, typically [0,1] or [ε,1+ε]. Its core idea is to eliminate the influence of data units by subtracting the minimum value and dividing by the difference between the maximum and minimum values, thus allowing all data points to be compared on the same scale. Here, x represents each original data point in the fused gas concentration sequence; these data points are the spliced ​​historical and predicted gas concentration values, which may have different absolute values. and These are the global minimum and maximum values ​​in the entire original fused gas concentration sequence. During normalization, all data points will be scaled relative to these two extreme values ​​to ensure that the data distribution of the entire sequence is mapped to the new target interval. ε is a very small positive number, its main function being to prevent [further distortion]. and This prevents calculation errors or program crashes when the denominator is zero (i.e., all data points in the sequence are identical). Furthermore, it ensures that the normalized value is not strictly zero, which provides numerical stability for subsequent steps involving ratio calculations (such as uncoded ratios).

[0046] This processing effectively eliminates the influence of differences in dimensions and numerical ranges between different gas concentration data, ensuring that all gas concentration data have equal weight and comparability in subsequent uncoded ratio calculations. For example, when calculating uncoded ratios such as C2H2 / C2H4, C2H4 / C2H6, and CH4 / H2, the normalized data ensures the stability and accuracy of the ratio calculation, avoiding ratio distortion caused by excessive differences in the magnitude of the original data. Simultaneously, the introduction of a very small positive number ε further enhances the numerical robustness of the normalization process, effectively avoiding potential division-by-zero errors. This provides a more reliable and consistent data foundation for subsequent ratio-based fault diagnosis, significantly improving the accuracy and reliability of power transformer fault diagnosis.

[0047] In one alternative implementation, the uncoded ratios include at least the C2H2 / C2H4 ratio r, the C2H4 / C2H6 ratio m, and the CH4 / H2 ratio k in the normalized fusion gas concentration sequence.

[0048] Simultaneously, fault diagnosis is performed according to preset time units to obtain the fault type time series results of the power transformer within the target life cycle, including: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] r , m , k The fault type of the power transformer is determined by comparing it with the preset threshold range for each time unit to determine whether it is in a low temperature overheating, medium temperature overheating, high temperature overheating, partial discharge, spark discharge or arc discharge in each time unit, thereby obtaining the fault type time sequence result of the power transformer within the target life cycle.

[0049] Specifically, uncoded ratios refer to indicators that reflect the characteristics of internal transformer faults by comparing specific gas concentrations. These ratios are commonly used in the Duval triangulation or its variations to identify different types of faults. In this embodiment, the ratios r of C2H2 / C2H4 and C2H4 / C2H6 are specifically selected. m and the CH4 / H2 ratio k These ratios, based on experience and theory from dissolved gas analysis (DGA) in power transformer oil, are widely considered effective indicators for diagnosing internal transformer faults (such as overheating and discharge). By normalizing the concentrations of these specific gases before calculating the ratios, the influence of differences in the absolute values ​​of gas concentrations between different transformers or at different times can be eliminated, making fault diagnosis more stable and reliable.

[0050] Fault diagnosis is performed according to preset time units to ensure the continuity and timeliness of fault diagnosis, so that a clear fault type determination can be obtained in each time unit (e.g., daily, weekly), thereby forming a complete fault type time sequence result. The specific diagnostic process involves applying the above calculations...r , m , k These three uncoded ratios are compared time-by-time with their respective pre-defined threshold ranges. These threshold ranges are determined based on extensive historical fault data, expert experience, and relevant industry standards (such as IEC and IEEE standards). For example, when... r , m , k When the ratio falls within a specific range, it may indicate low-temperature overheating; when it falls within another range, it may indicate partial discharge. Through this time-unit-by-time comparison, the system can dynamically identify the fault type of the power transformer in each time unit, including one of the following: low-temperature overheating, medium-temperature overheating, high-temperature overheating, partial discharge, spark discharge, or arc discharge. This detailed diagnosis helps to more accurately capture the evolution of the fault.

[0051] In this embodiment, the no-coding ratio judgment method is shown in Table 1.

[0052] Table 1

[0053] In one optional implementation, steps S6 and S7 specifically include: If adjacent time units have the same fault type, they are merged into the same consecutive fault segment. When the normal interval between two similar fault segments does not exceed the preset recovery threshold... τ Bridged merging was performed to obtain time series containing various fault types. The fault types in the time series were then mapped to Class A and Class B faults, resulting in time series containing only Class A and Class B faults. Class A faults include low-temperature overheating, medium-temperature overheating, partial discharge, and spark discharge; Class B faults include high-temperature overheating and arc discharge. The number of days that Class A faults occurred was counted. D A Number of days of Class B faults D B Number of Class A failure events F A Number of Class B failure events F B and the maximum cumulative duration of Class A faults L A and the maximum cumulative duration of Class B faults L B .

[0054] Specifically, when performing state transition analysis on fault type time-series results, firstly, when adjacent time units have the same fault type, they are merged into the same continuous fault segment. This aims to ensure that a single, uninterrupted fault, even if detected across multiple consecutive time units, is considered a continuous event. For example, if a "low temperature overheating" fault is diagnosed for two consecutive days, it is considered a two-day "low temperature overheating" event, rather than two independent fault events, thus avoiding duplicate counting of persistent faults. Secondly, when the normal interval between two similar fault segments does not exceed a preset recovery threshold... τ When bridging and merging fault segments, this mechanism is specifically designed to handle intermittent faults. If the same type of fault reappears after a brief "normal" period (i.e., a normal period shorter than a preset recovery threshold τ), this usually indicates that the previous fault was not completely eliminated but only temporarily alleviated. By bridging and merging these fault segments, this method can more accurately capture the true cumulative duration and frequency of such recurring problems, avoiding underestimating the actual impact of a persistent problem. Preset recovery threshold τ The configuration allows the system to tolerate short-term fault recovery illusions to a certain extent, thus more realistically reflecting the persistence and recurrence of faults. Through the above merging and bridging operations, a time series containing various fault types is finally obtained, which can more accurately reflect the dynamic process of fault occurrence, persistence, and recovery of power transformers within the target life cycle.

[0055] Based on this, the fault types in the time series containing various fault types are mapped to Class A and Class B faults, thus obtaining time series containing only Class A and Class B faults. This step clearly defines the classification criteria for fault severity. Class A faults are defined as low-temperature overheating, medium-temperature overheating, partial discharge, and spark discharge. These faults typically represent early or mid-stage degradation of power transformers and can be resolved through routine maintenance or partial repairs; their impact on equipment operation and the required maintenance costs are relatively low. Class B faults, on the other hand, are defined as high-temperature overheating and arc discharge. These faults typically represent more severe degradation, such as insulation breakdown or severe overheating, which may lead to emergency shutdown or even scrapping of equipment; their repair or replacement costs are significantly higher, and they pose a greater risk to grid operation. This clear classification lays a solid foundation for subsequent differentiated cost accounting, allowing faults of different severity to be assigned different economic weights.

[0056] Subsequently, statistics were compiled for the mapped Class A and Class B faults, specifically including: counting the number of days Class A faults occurred. D A Number of days of Class B faults D B Number of Class A failure events FA Number of Class B failure events F B and the maximum cumulative duration of Class A faults L A and the maximum cumulative duration of Class B faults L B These specific statistical indicators provide a multi-dimensional quantification of fault characteristics. Among them, the number of days incurred ( D A , D B The number of fault events reflects the total duration of the fault throughout the entire target lifespan, measuring the overall breadth of the fault's impact on equipment operation; F A , F B Statistics on the occurrence frequency of independent failure events help assess the recurrence and potential root causes of failures; maximum cumulative duration ( L A , L B This captures the duration of a single, most severe, or most persistent failure event, which is particularly important for assessing the risks and costs under extreme failure conditions.

[0057] Through the aforementioned technical solutions, this application effectively addresses the problem of ambiguous fault event boundaries by accurately identifying, merging, and bridging fault events, ensuring the accuracy of fault duration and frequency. Simultaneously, the clearly defined Class A and Class B fault classification standards enable objective and consistent assessment of faults of varying severity, providing a reliable basis for subsequent cost calculations. Furthermore, multi-dimensional statistical indicators (number of days, number of events, and maximum cumulative duration) can more comprehensively and precisely characterize the severity and impact of faults, providing solid data support for the accurate calculation of subsequent dynamic fault costs. These improvements collectively enhance the accuracy and practicality of power transformer lifecycle cost prediction, enabling the prediction results to more realistically reflect the potential risks and economic losses during equipment operation.

[0058] In one optional implementation, a preset recovery threshold is set. τ The monitoring time granularity is set to one or more time units to suppress the interference of short-term recovery fluctuations on fault event segmentation.

[0059] Specifically, preset recovery threshold τThis is a key parameter used to determine whether bridging and merging should be performed between two faulty segments of the same type. It defines the maximum permissible "normal" interval between two faulty segments when the fault types are the same. If the actual normal interval is less than or equal to this threshold, the two faulty segments are considered to be continuations of the same fault event and should be merged. Monitoring time granularity refers to the temporal resolution of data acquisition and fault diagnosis, such as one day, half a day, or several hours. A preset recovery threshold is set. τ Setting it to one or more time units means that the threshold is directly related to the actual monitoring frequency. For example, if the monitoring time granularity is one day, then... τ It can be set to 1 day, 2 days, etc. This setting ensures that the threshold matches the time characteristics of the actual data, avoiding the incompatibility that may arise from using a fixed time value. In actual operation, power transformers may experience brief, non-substantial "recovery" phenomena, such as a gas concentration briefly decreasing and then rapidly rising again. If the recovery threshold is set too low or unreasonable, these short-term fluctuations may cause a continuous fault event to be incorrectly divided into multiple independent fault events. By... τ Setting one or more time units that match the monitoring time granularity can effectively "tolerate" these short-term fluctuations, avoid their incorrect segmentation of fault events, and thus more accurately identify and count actual fault events.

[0060] In one alternative implementation, the dynamic failure cost is expressed as: ; ; ; in, This represents the total dynamic failure cost. and The dynamic failure costs are for Class A and Class B failures, respectively. , These represent the number of days that Class A and Class B faults occurred within the target lifespan, respectively. , These are the number of fault events for Class A faults and Class B faults, respectively. , These are the maximum cumulative durations of Class A and Class B faults, respectively. , The average repair costs for Class A and Class B faults are respectively. , , These are the weighting coefficients corresponding to the number of days a Class A fault occurs, the number of fault events, and the maximum cumulative duration, respectively. , , These are the weighting coefficients corresponding to the number of days of occurrence of Category B faults, the number of fault events, and the maximum cumulative duration, respectively.

[0061] Dynamic Failure Cost This refers to the total economic loss incurred by a power transformer due to various faults within its target lifespan. This dynamic fault cost divides the total cost into Class A and Class B faults. and This allows for differentiated cost assessment and management of faults of varying severity, thereby more accurately reflecting the actual impact of different fault types on equipment operation and maintenance.

[0062] Through the aforementioned technical solution, this application can effectively transform information such as the number of days with faults, the frequency of fault events, and the maximum cumulative duration of faults, statistically obtained from the statistics of power transformers within their target lifespan, into quantified dynamic fault costs through a structured cost calculation model. This calculation method not only distinguishes between the different severity levels of Class A and Class B faults but also assigns them independent cost calculation logic and weighting coefficients, making cost assessment more refined and accurate. Furthermore, by introducing average maintenance costs and adjustable weighting coefficients, this solution can flexibly adapt to different operating environments and maintenance strategies, thereby more comprehensively and dynamically reflecting the actual impact of faults on the entire life-cycle cost of power transformers, providing a more reliable economic basis for equipment management and maintenance decisions.

[0063] In addition, such as Figure 3 As shown, the present invention also provides a power transformer life-cycle cost prediction system for implementing the above method, comprising: The data acquisition module is used to collect multi-source data of the power transformer during operation; the multi-source data includes: historical gas concentration sequence data of various target gases dissolved in oil and full life cycle cost parameters; The data preprocessing module is used to preprocess the historical gas concentration sequence data of each target gas, and to perform CEEMDAN decomposition and reconstruction on the preprocessed historical gas concentration sequence data to obtain the reconstructed input sequence of each target gas. The decomposition prediction module is used to input the reconstructed input sequences of each target gas into the LSTM model to obtain the predicted gas concentration sequence within the future target prediction period; The fusion sequence construction module is used to splice historical gas concentration sequences and predicted gas concentration sequences according to a unified time axis to form a fusion gas concentration sequence covering the target life cycle, and to perform Min-Max normalization on the fusion gas concentration sequence. The fault diagnosis module is used to calculate the uncoded ratio based on the fused gas concentration sequence after Min-Max normalization, and to perform fault diagnosis according to the preset time unit to obtain the fault type time series results of the power transformer within the target life cycle. The State Transition Analysis and Event Construction Module is used to perform state transition analysis on the time sequence results of fault types, merge the time units of adjacent faults of the same type into continuous fault segments, and perform bridging and merging on the time units of faults of the same type with an interval not exceeding a preset recovery threshold to obtain a fault event sequence. The severity mapping module is used to map the fault types in the fault event sequence to Class A faults and Class B faults, and to count the number of days, frequency of fault events and maximum cumulative duration of Class A and Class B faults in the fault event sequence, respectively. The failure cost calculation module is used to calculate dynamic failure costs based on the statistical results output by the severity mapping module. The full life cycle cost output module is used to summarize the dynamic fault cost and the initial total cost to output the full life cycle cost of the power transformer.

[0064] The following example will provide a more detailed explanation of the above technical solution: This embodiment uses a 110kV / 50MVA power transformer as an example, with a target lifespan of 12 years. Historical concentration data of dissolved gases in the oil (H2, CH4, C2H2, C2H4, C2H6, CO, and CO2, with a monitoring period of 1 day) have been collected for the first two years (2023-2025) after the transformer was put into operation. The goal is to predict the total lifespan cost for the next 10 years (2026-2035).

[0065] Step S1: Collect data from multiple sources.

[0066] The collection of multi-source data includes: Historical concentration sequences of seven characteristic gases (3652 time points); Life cycle cost parameters: Investment cost of RMB 1.37 million (equipment purchase of RMB 1.25 million + installation of RMB 120,000), average annual operating cost of RMB 20,000 (loss of RMB 16,000 + switching of RMB 4,000), average annual maintenance cost of RMB 9,000, decommissioning cost of RMB 60,000, average repair cost of Class A faults of RMB 10,000 / time, and average repair cost of Class B faults of RMB 75,000 / time.

[0067] Step S2: Preprocessing and CEEMDAN decomposition and reconstruction.

[0068] The historical gas concentration sequences are processed using methods including missing value completion (linear interpolation), outlier removal (3σ criterion), and time alignment. Taking C2H4 as an example, CEEMDAN decomposition is performed, yielding multiple intrinsic mode function (IMF) components and one residual component. After removing the first three high-frequency noise components, the remaining IMF components are superimposed with the residual component to obtain the reconstructed input sequence. The other six gases are processed in the same way.

[0069] Step S3: LSTM prediction.

[0070] The reconstructed gas sequences were input into an LSTM model (30-day window, two LSTM layers, 64 units each) to predict daily gas concentrations over the next 10 years (2026-2035). The LSTM predictions for each IMF component and the residual component were then superimposed to obtain a predicted sequence of 3652 time points.

[0071] Step S4: Constructing the fusion sequence and Min-Max normalization.

[0072] The historical sequence (731 points) and the predicted sequence (3652 points) were concatenated along the time axis to form a fused sequence covering 12 years (4383 points in total). Min-Max normalization was performed for each gas. For example, C2H4 , , =10 -8 .

[0073] Step S5: Fault diagnosis.

[0074] Calculate three uncoded ratios daily: r =C2H2 / C2H4, m =C2H4 / C2H6, k =CH4 / H2. Compare with a preset threshold to determine the fault type (low-temperature overheating, medium-temperature overheating, high-temperature overheating, partial discharge, spark discharge, arc discharge). For example, on a certain day... r =0.05, m =2.1, k =0.6, which is determined to be medium-temperature overheating. Daily diagnosis yielded 12 years of fault type timing results.

[0075] Step S6: Fault event construction.

[0076] Set recovery threshold τ =3 days. Adjacent fault time units of the same type are merged into a continuous fault segment; if the normal interval between two fault segments of the same type is ≤3 days, they are bridged and merged into the same fault event. For example: low temperature overheating occurs for 2 days, normal for 2 days, and then occurs for another 2 days, which are merged into a single event lasting 6 days. The final fault event sequence is obtained.

[0077] Step S7: Severity mapping and statistics.

[0078] Map the fault type to: Class A faults (minor): low temperature overheating, medium temperature overheating, partial discharge, spark discharge; Class B faults (serious): overheating, arc discharge.

[0079] Statistics over 12 years: Category A: Number of days incurred D A =320 days, number of failure events F A =15 times, maximum cumulative duration L A =38 days; - Category B: Number of days incurred D B =35 days, number of failure events F B =3 times, maximum cumulative duration L B =18 days.

[0080] Step S8: Dynamic failure cost calculation.

[0081] Set weighting coefficients: Class A =0.2, =0.4, =0.4; Class B =0.1, =0.5, =0.4, the above weighting coefficients can be obtained by performing multiple linear regression analysis on historical fault data.

[0082] calculate: Ten thousand yuan; Ten thousand yuan; Ten thousand yuan.

[0083] Step S9: As Figure 4 As shown, this is a summary of the total lifecycle cost.

[0084] Investment cost: 1.37 million yuan; Operating costs: 12 years × 2.0 = 240,000 yuan; Maintenance cost: 12 years × 0.9 = 108,000 yuan; Retirement cost: 60,000 yuan; Dynamic failure cost: RMB 1.767 million; The total cost is 137 + 24 + 10.8 + 6 + 176.7 = 354.5 million yuan, of which dynamic failure cost accounts for 49.8%, making it one of the main cost items.

[0085] This embodiment fully demonstrates the entire process from data acquisition, CEEMDAN-LSTM prediction, fusion diagnosis, fault event construction to dynamic fault cost calculation, and verifies that the method can effectively predict the full life cycle cost of power transformers, providing a quantitative basis for operation and maintenance decisions.

[0086] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for predicting the full life-cycle cost of a power transformer, characterized in that, Includes the following steps: S1. Collect multi-source data of the power transformer during operation; the multi-source data includes: historical gas concentration sequence data of various target gases dissolved in oil and full life cycle cost parameters; S2. Preprocess the historical gas concentration sequence data of each target gas, and perform CEEMDAN decomposition and reconstruction on the preprocessed historical gas concentration sequence data to obtain the reconstructed input sequence of each target gas. S3. Input the reconstructed input sequences of each target gas into the LSTM model to obtain the predicted gas concentration sequence within the future target prediction period. S4. The historical gas concentration sequence and the predicted gas concentration sequence are spliced ​​together according to a unified time axis to form a fused gas concentration sequence covering the target life cycle, and the fused gas concentration sequence is subjected to Min-Max normalization. S5. Calculate the uncoded ratio based on the fused gas concentration sequence after Min-Max normalization, and perform fault diagnosis according to the preset time unit to obtain the fault type time series results of the power transformer within the target life cycle. S6. Perform state transition analysis on the time sequence results of the fault type, merge the time units of adjacent similar faults into continuous fault segments, and perform bridging and merging on the time units of similar faults with an interval not exceeding a preset recovery threshold to obtain a fault event sequence. S7. Map the fault types in the fault event sequence to type A faults and type B faults, and count the number of days, frequency of fault events and maximum cumulative duration of type A and type B faults in the fault event sequence respectively. S8. Calculate the dynamic failure cost based on the statistical results of S7; S9. Summarize the dynamic fault cost with the initial total cost to output the full life cycle cost of the power transformer.

2. The method according to claim 1, characterized in that, The target gases include: H2, CH4, C2H2, C2H4, C2H6, CO, and CO2; The full life cycle cost parameters include: investment cost, operating cost, maintenance cost, decommissioning cost, average repair cost for Class A faults, and average repair cost for Class B faults; wherein, the investment cost includes equipment purchase cost and installation cost; the operating cost includes operating loss cost and switching cost.

3. The method according to claim 1, characterized in that, The preprocessing includes missing value completion, outlier removal, and time alignment.

4. The method according to claim 1, characterized in that, The CEEMDAN decomposition and reconstruction includes: Historical gas concentration sequence data of each target gas after preprocessing Perform CEEMDAN decomposition separately to obtain N Vectors of intrinsic mode function components arranged in descending order of frequency. and a residual component vector Its expression is: in, For the first t The target gas concentration values ​​corresponding to each time point ; For the first i One intrinsic mode function component; In order to be in t Historical concentration sequence data at each time point The middle belongs to the first i The amplitude of each intrinsic mode function component; The residual component; exist t Historical concentration sequence data at each time point The sum of the remaining components after decomposition; The total number of intrinsic mode function components obtained after decomposition; This represents the total number of historical time points. Identify and remove the preceding characters that represent high-frequency random noise. k One intrinsic mode function component , the remaining The low-to-mid-frequency intrinsic mode function components are superimposed with the residual components to obtain the reconstructed input sequences for each target gas. , of which t Reconstruction values ​​at each time point The expression is: 。 5. The method according to claim 1, characterized in that, The Min-Max normalization process is expressed as follows: in, The raw fusion gas concentration sequence data to be normalized. and These are the minimum and maximum values ​​in the original fusion gas concentration sequence, respectively. To prevent the use of a preset minimum positive number with a denominator of 0.

6. The method according to claim 2, characterized in that, The uncoded ratio includes at least the C2H2 / C2H4 ratio in the normalized fusion gas concentration sequence. r The ratio of C2H4 / C2H6 m and the CH4 / H2 ratio k ; The step of performing fault diagnosis according to a preset time unit to obtain the fault type time sequence results of the power transformer within the target life cycle includes: [The sentence is incomplete and requires further context to be translated accurately.] r , m , k The fault type of the power transformer is determined by comparing it with the preset threshold range for each time unit to determine whether it is in a low temperature overheating, medium temperature overheating, high temperature overheating, partial discharge, spark discharge or arc discharge in each time unit, so as to obtain the fault type time sequence result of the power transformer within the target life cycle.

7. The method according to claim 6, characterized in that, Specifically, S6 and S7 include: When adjacent time units have the same fault type, they are merged into the same continuous fault segment. When the normal interval between two similar fault segments does not exceed the preset recovery threshold... τ Bridge merging is performed to obtain time series containing each fault type; The fault types in the time series containing various fault types are mapped to Class A faults and Class B faults, resulting in a time series containing only Class A faults and Class B faults; wherein, Class A faults include low temperature overheating, medium temperature overheating, partial discharge and spark discharge, and Class B faults include high temperature overheating and arc discharge. Count the number of days that Category A faults occurred. D A Number of days of Class B faults D B Number of Class A failure events F A Number of Class B failure events F B and the maximum cumulative duration of Class A faults L A and the maximum cumulative duration of Class B faults L B .

8. The method according to claim 7, characterized in that, The preset recovery threshold τ The monitoring time granularity is set to one or more time units to suppress the interference of short-term recovery fluctuations on fault event segmentation.

9. The method according to claim 7, characterized in that, Dynamic failure cost is expressed as: ； ； ； in, This represents the total dynamic failure cost. and The dynamic failure costs are for Class A and Class B failures, respectively. , These represent the number of days that Class A and Class B faults occurred within the target lifespan, respectively. , These are the number of fault events for Class A faults and Class B faults, respectively. , These are the maximum cumulative durations of Class A and Class B faults, respectively. , The average repair costs for Class A and Class B faults are respectively. , , These are the weighting coefficients corresponding to the number of days a Class A fault occurs, the number of fault events, and the maximum cumulative duration, respectively. , , These are the weighting coefficients corresponding to the number of days of occurrence of Category B faults, the number of fault events, and the maximum cumulative duration, respectively.

10. A power transformer life-cycle cost prediction system, characterized in that, For implementing the method of any one of claims 1 to 9, comprising: The data acquisition module is used to collect multi-source data of the power transformer during operation; the multi-source data includes: historical gas concentration sequence data of various target gases dissolved in oil and full life cycle cost parameters; The data preprocessing module is used to preprocess the historical gas concentration sequence data of each target gas, and to perform CEEMDAN decomposition and reconstruction on the preprocessed historical gas concentration sequence data to obtain the reconstructed input sequence of each target gas. The decomposition prediction module is used to input the reconstructed input sequences of each target gas into the LSTM model to obtain the predicted gas concentration sequence within the future target prediction period; The fusion sequence construction module is used to splice historical gas concentration sequences and predicted gas concentration sequences according to a unified time axis to form a fusion gas concentration sequence covering the target life cycle, and to perform Min-Max normalization processing on the fusion gas concentration sequence. The fault diagnosis module is used to calculate the uncoded ratio based on the fused gas concentration sequence after Min-Max normalization, and to perform fault diagnosis according to the preset time unit to obtain the fault type time series results of the power transformer within the target life cycle. The state transition analysis and event construction module is used to perform state transition analysis on the time series results of the fault type, merge the time units of adjacent similar faults into continuous fault segments, and perform bridging and merging on the time units of similar faults with an interval not exceeding a preset recovery threshold to obtain a fault event sequence. The severity mapping module is used to map the fault types in the fault event sequence to Class A faults and Class B faults, and to count the number of days, frequency of fault events and maximum cumulative duration of Class A and Class B faults in the fault event sequence, respectively. The failure cost calculation module is used to calculate dynamic failure costs based on the statistical results output by the severity mapping module. The full life cycle cost output module is used to summarize the dynamic fault cost and the initial total cost, and output the full life cycle cost of the power transformer.

Images