Post-disaster main transformer equipment loss degree quantitative evaluation method

Abstract

The invention provides a post-disaster main transformer equipment loss degree quantitative evaluation method, which comprises the steps of firstly obtaining a natural disaster category and main transformer equipment fault data, constructing a disaster-caused fault influence factor system, and calculating a disaster weight and a fault mode weight; listing and preprocessing equipment maintenance evaluation indexes based on historical disaster damage data and equipment operation states, identifying a core evaluation index set, and calculating importance weights of the core indexes; establishing a loss index calculation model according to the weights, calculating a basic loss index of the fault sample data, and presetting a loss grade interval; then carrying out nonlinear calibration on the basic loss index, and fusing multi-dimensional features to generate a comprehensive loss index; and finally, optimizing the comprehensive loss index through a random forest regression model, outputting a final loss prediction index, and judging the equipment repair priority according to the final loss prediction index. According to the evaluation method, unified and quantifiable loss indexes and levels are output, and a decision basis is provided for post-disaster repair priority judgment and manpower and material resource scheduling.

Description

Technical Field

[0001] This application relates to the field of industrial equipment damage assessment, and in particular to a method for quantitatively assessing the degree of loss of main transformer equipment after a disaster. Background Technology

[0002] As a crucial component of the power grid, the main transformer plays a vital role in ensuring the safe and stable operation of the power system and the normal production and lives of the people. Extreme natural disasters such as typhoons, blizzards, floods, and thunderstorms often damage main transformer equipment, causing widespread power outages and even power system collapse, threatening people's lives, business operations, and social stability. Rapid and large-scale quantitative assessment of the extent of main transformer damage after a natural disaster is essential for conducting centralized overhauls and restoring power supply as quickly as possible. This is not only a key aspect of power system emergency response but also a fundamental guarantee for post-disaster equipment repair.

[0003] Currently, the assessment of the extent of damage to main transformer equipment relies heavily on existing loss experience, resulting in low overall efficiency. Furthermore, natural disasters within the jurisdiction of provincial units exhibit diverse characteristics, lacking a targeted assessment model for main transformer equipment failures caused by specific regions or disasters. Simultaneously, during major overhauls of main transformer equipment, issues such as uneven resource allocation, unclear priorities for repairing different types of faults, and a lack of clear standards for personnel and material scheduling may arise, affecting the overall progress of equipment repair and increasing overhaul costs.

[0004] In conclusion, there is an urgent need to develop a scientific method for assessing the extent of damage to main transformer equipment after a disaster. This method would enable rapid assessment of equipment damage status, determination of repair priorities, and dynamic allocation of manpower and resources. This would allow for efficient completion of major equipment overhauls, maximizing resource utilization efficiency, and enhancing the overall effectiveness and resilience of the power system in post-disaster recovery. Summary of the Invention

[0005] To address the issues of current methods for assessing damage to main transformer equipment being too subjective, relatively inefficient, and lacking sufficient quantification of the extent of damage, leading to uneven allocation of maintenance resources in the later stages.

[0006] This application provides a method for quantitatively assessing the degree of damage to main transformer equipment after a disaster, including:

[0007] Acquire natural disaster category data and main transformer equipment failure data, construct a disaster-induced failure influencing factor system, and calculate disaster weights and failure mode weights;

[0008] Based on historical disaster damage data and equipment operating status, equipment maintenance assessment indicators are listed, data preprocessing is performed on the equipment maintenance assessment indicators, a core assessment indicator set is identified, and the importance weight of the core indicators is calculated.

[0009] Based on the disaster weight, the failure mode weight, and the importance weight of the core indicators, a loss index calculation model is established to calculate the basic loss index for failure sample data and preset the loss level range.

[0010] The basic loss index is nonlinearly calibrated, and multidimensional features are fused to generate a comprehensive loss index;

[0011] The comprehensive loss index is optimized using a random forest regression model to output a final loss prediction index, and the equipment repair priority is determined based on the final loss prediction index.

[0012] In one feasible implementation, the specific steps for constructing the disaster-induced failure influencing factor system include:

[0013] Natural disasters are categorized and coded to form a disaster category hierarchy;

[0014] For each of the aforementioned natural disaster categories, a corresponding failure mode is matched to form a failure mode layer;

[0015] A disaster damage severity comparison matrix and a failure damage probability matrix were constructed based on expert scoring data.

[0016] The weights for each of the natural disaster categories are calculated using geometric averaging and normalization.

[0017] Weights are calculated for each failure mode under each of the aforementioned natural disaster categories, and a consistency check is performed.

[0018] Output a global fault weight ranking, which is used to identify the repair priority of the device.

[0019] In one feasible implementation, the specific steps for identifying the core evaluation index set include:

[0020] Integrate major equipment maintenance indicators, fault correlation indicators, and basic equipment attribute indicators to form an original indicator matrix;

[0021] The original index matrix is ​​processed for outliers and made dimensionless to generate standardized data;

[0022] By using covariance analysis and eigenvalue calculation, core features representing the preset amount of information are selected.

[0023] The correlation between the core features and the degree of fault damage is learned using a neural network model, and the importance weights of the core indicators are calculated.

[0024] The core indicators are then further filtered based on importance thresholds, and the set of core evaluation indicators and their corresponding weights are output.

[0025] In one feasible implementation, the specific steps for calculating the basic loss index include:

[0026] Based on power grid maintenance guidelines and historical statistical characteristics, preset boundary conditions for loss levels are established.

[0027] By combining the importance weights of the core indicators and the failure mode weights, the indicator-layer weighted score of each failure sample is calculated.

[0028] The fault layer weighted score is calculated based on the weighted score of the indicator layer and the fault concurrency coefficient.

[0029] The weighted scores of all the fault layers are normalized and mapped to a preset interval to generate the basic loss index.

[0030] In one feasible implementation, the specific steps for nonlinearly calibrating the basic loss index include:

[0031] The core indicators are categorized into threshold indicators, linear indicators, and saturation indicators.

[0032] Design a corresponding calibration function for each type of core indicator, and apply the calibration function to the core indicator score;

[0033] The weighted score of the calibrated index layer is recalculated and combined with the failure mode weights to generate a calibration loss index;

[0034] The comprehensive loss index is generated by integrating the basic loss index, the calibration loss index, and the disaster-related weights.

[0035] In one feasible implementation, the specific steps for optimizing the comprehensive loss index include:

[0036] Construct a multidimensional correlation matrix that includes core loss features, weight features, and device scenario features;

[0037] Outlier correction is performed on the comprehensive loss index to ensure the rationality of data distribution;

[0038] We employ parallel training of the Random Forest model and the XGBoost model, and dynamically adjust the model parameters.

[0039] Based on the validation set performance and weight allocation, calculate the model fusion weights;

[0040] The final loss prediction index is output by weighted average.

[0041] In one feasible implementation, the outlier handling uses the 3σ principle to identify outliers and replaces them with the median;

[0042] The dimensionless processing employs the Z-Score method to achieve data standardization.

[0043] In one feasible implementation, the loss level boundary conditions are dynamically calculated based on the index score quantiles, including: index ranges for minor loss, moderate loss, and severe loss.

[0044] In one feasible implementation, the calibration function includes: an exponential function, a hyperbolic tangent function, and a saturation function;

[0045] The exponential function is used for the threshold-type indicator;

[0046] The hyperbolic tangent function is used for the linear index;

[0047] The saturation function is used for the saturation-type index.

[0048] In one feasible implementation, the parameters of the random forest regression model are dynamically set according to the fusion weights and disaster resistance level, including: the number of decision trees, the maximum depth, and the learning rate;

[0049] The constraint for training the random forest regression model is to ensure that the predicted level matches the initial level.

[0050] This application presents a quantitative assessment method for the degree of damage to main transformer equipment after a disaster. It utilizes the analytic hierarchy process (AHP) to systematically transform the qualitative experience of domain experts into quantitative weights, clarifying the destructive contribution of different disaster-induced failure modes and effectively addressing the issues of strong subjectivity and reliance on personal experience in traditional assessments. Furthermore, it employs principal component analysis (PCA) to reduce the dimensionality of massive, multi-dimensional fault monitoring indicators, extracting the most representative core assessment dimensions and overcoming the shortcomings of redundant indicators and ambiguous assessment focus in traditional methods. Further, it introduces a random forest regression model to integrate disaster weights and core indicators, effectively characterizing the complex nonlinear relationship between disaster, failure, and the degree of main transformer loss, significantly improving the accuracy of loss index calculation and model adaptability. This comprehensive assessment method ultimately outputs a unified and quantifiable loss index and level, providing a clear and reliable decision-making basis for post-disaster repair priority determination and dynamic scheduling of human and material resources, thereby achieving optimal allocation of maintenance resources and improving the overall efficiency and resilience of power system post-disaster recovery. Attached Figure Description

[0051] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the implementation of the invention and, together with the description, serve to explain the principles of the embodiments of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0052] Figure 1 This is a flowchart illustrating a method for quantitatively assessing the degree of damage to main transformer equipment after a disaster, as shown in one embodiment of this application. Detailed Implementation

[0053] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the embodiments of the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of how embodiments of the invention are carried out.

[0054] This application aims to establish a method for assessing the extent of damage to main transformer equipment after a disaster, in order to address the problems of current methods for assessing main transformer equipment disaster damage being highly subjective, relatively inefficient, and lacking sufficient quantification of the extent of damage, leading to uneven allocation of subsequent maintenance resources. (Refer to...) Figure 1 As shown, the method includes the following steps:

[0055] S100: Obtain natural disaster category data and main transformer equipment fault data, construct a disaster-induced fault influencing factor system, and calculate disaster weights and fault mode weights.

[0056] The natural disaster category data includes common disaster types such as typhoons, floods, thunderstorms, and blizzards. These disasters are classified through coding methods; for example, floods correspond to fault modes such as water immersion in the equipment itself and insulation oil contamination. Fault data is derived from historical disaster damage records and equipment operation logs.

[0057] In some embodiments, the specific steps further include:

[0058] S110: Encode natural disaster categories to form a disaster category layer.

[0059] The natural disaster category coding uses floods, typhoons, lightning strikes, and blizzards as the criteria layer. Each disaster corresponds to a specific failure mode, such as floods causing the main body and accessories to be submerged in water and become damp, or insulation oil contamination and deterioration. The role of the disaster category layer is to classify disaster types and provide a structural framework for weight calculation.

[0060] S120: For each type of natural disaster, a corresponding failure mode is matched to form a failure mode layer.

[0061] Specifically, when constructing the influencing factor system, the analytic hierarchy process (AHP) is adopted, with natural disaster categories as the criterion layer and failure modes as the indicator layer, forming a three-layer structure. The main objective is to accurately assess the contribution of various natural disasters to the losses of the main transformer equipment caused by different failure modes (i.e., the A objective layer). First, various common natural disasters are coded according to their categories (i.e., the B criterion layer), and then a failure influencing factor indicator system is designed (i.e., the C indicator layer), as shown in Table 1 below:

[0062] Table 1 Specific Fault Categories of Main Transformer Equipment

[0063]

[0064] S130: Construct a disaster damage degree comparison matrix and a failure damage probability matrix based on expert scoring data.

[0065] Table 1 above lists the types of faults that various natural disasters may cause. Since different natural disasters have different natures and varying degrees of overall damage, to identify which fault has the highest probability of occurrence and causes the greatest damage under different types of natural disasters, it is first necessary to construct a comprehensive damage matrix for each type of disaster, i.e., which disaster causes the strongest damage to the power grid compared to any other disaster (represented by matrix A); secondly, to construct a probability matrix for faults occurring under specific disaster conditions, i.e., which fault has the highest relative probability of occurrence under the influence of a specific disaster (represented by matrix B), as shown in Table 2 below:

[0066] Table 2A Matrix

[0067]

[0068]

[0069] As shown above, where a ij The numerical values ​​of (i∈(1,4),j∈(1,4)) represent the overall destructive degree of disaster i (i∈(1,4)) relative to disaster j (j∈(1,4)), let a ij ∈(1 / 4,4) and is a non-zero integer. When its value is 4, it means that disaster i (i∈(1,4)) is extremely destructive and has a significant difference in intensity compared to disaster j (j∈(1,4)). When its value is 1 / 4, it means that disaster i (i∈(1,4)) is significantly weaker than disaster j (j∈(1,4)). When its value is 1, it means that the two natural disasters have similar overall damage to the main transformer equipment.

[0070] S140: Calculate the weight of each natural disaster category through geometric mean and normalization.

[0071] After obtaining the expert score for each line, calculate their geometric mean using the following formula:

[0072]

[0073] Normalizing the geometric mean values ​​yields the weights of each disaster, which represent their average degree of damage.

[0074]

[0075] The next step is to construct a fault judgment matrix using a specific natural disaster as an example, and study the relative magnitude of the damage caused by a specific fault under this type of disaster:

[0076] Table 3 C matrix

[0077]

[0078] As shown in the table above, c ij ∈(1 / 9,9),c ij The value represents the degree of damage of fault i relative to fault j within this matrix (i.e., from the perspective of a specific natural disaster). When c ij When c = 1 / 9, it indicates that fault i is the least damaging relative to fault j; when c ij When the value is 9, it indicates that the damage caused by fault i is much greater than that caused by fault j.

[0079] S150: Calculate the weights of failure modes under each natural disaster category and perform a consistency check.

[0080] The calculation process for the consistency check of all the above matrices is as follows:

[0081] S151: Normalize the elements in each column and construct a normalized matrix:

[0082]

[0083] S152: Calculate the eigenvector w using the sum-product method;

[0084] S153: Calculate the product vector: DW = D * W;

[0085] S154: Calculate the maximum eigenvalue of each matrix:

[0086]

[0087] S155: Calculate the consistency index: In this scenario: n∈(1,9);

[0088] S156: Calculate the consistency ratio: CR < 0.10;

[0089] S157: Calculate the importance of each fault indicator to determine the priority of equipment overhaul. Let θ represent the global weight of a certain type of fault. Taking type 1 fault (the main body and accessories are submerged in water and damp) as an example, the calculation formula is as follows:

[0090] θ1=θ B1 *θ B1-C1 +θ B2 *θ B2-C1 +...+θ B9 *θ B9-C1 ;

[0091] Where θ i This represents the natural disaster weights that have been calculated in the B criterion layer.

[0092] S160: Output global fault weight sorting, which is used to identify the priority of equipment repair.

[0093] The calculation model for the damage weights of the other eight fault categories is similar. After all fault indicators have been calculated, the final weight calculation results are output and sorted. The fault category with the highest weight is the one that should be given priority when considering damage.

[0094] The purpose of this step is to transform qualitative experience into quantitative data, clarify the contribution of different disasters to equipment damage, lay the foundation for subsequent assessments, and reduce subjective judgments through systematic weight calculations, thereby improving the objectivity and repeatability of the assessment.

[0095] S200: Based on historical disaster damage data and equipment operating status, list equipment maintenance assessment indicators, perform data preprocessing on the equipment maintenance assessment indicators, identify the core assessment indicator set, and calculate the importance weight of the core indicators.

[0096] The equipment maintenance assessment indicators include major categories such as comprehensive transformer maintenance, accessory maintenance, and drying treatment. Each major category is further subdivided into specific indicators such as winding DC resistance imbalance rate and absorption ratio. Additionally, fault correlation indicators (such as fault concurrency coefficient) and basic equipment attribute indicators (such as service life) are supplemented.

[0097] The identification of core indicators is performed using principal component analysis (PCA). This involves calculating the covariance matrix and eigenvalues, selecting core features that represent over 90% of the information, and then using a feedforward neural network model to learn the correlation between these features and the degree of fault damage, calculating their importance weights. For example, the neural network uses ReLU and Sigmoid activation functions, optimizing parameters through training to output a set of core indicators.

[0098] The purpose of this step is to reduce dimensionality and focus on key indicators, avoid indicator redundancy, and improve evaluation efficiency. The key to solving the technical problem lies in extracting representative indicators through a data-driven approach, overcoming the shortcomings of traditional methods that suffer from vague focus.

[0099] In some embodiments, the specific steps for identifying the core evaluation metric set include:

[0100] S210: Integrate major equipment maintenance indicators, fault correlation indicators, and basic equipment attribute indicators to form an original indicator matrix;

[0101] Specifically, the assessment of main transformer equipment maintenance is matched with the seven major equipment maintenance categories according to the quota, and these assessment indicators are defined and integrated into basic maintenance indicators. On this basis, fault correlation indicators and basic equipment attribute indicators are added to form three types of indicator sets. At the same time, relevant data from the original cases are extracted to form the original indicator matrix, as shown in Table 4 below.

[0102] Table 4 Corresponding Indicators

[0103]

[0104]

[0105] S220: Perform outlier processing and dimensionless transformation on the original index matrix to generate standardized data.

[0106] Specifically, one-hot encoding will be used for the two classification features: fault propagation risk level Rs and disaster resistance level Lr. The encoding dimensions are determined according to the total number of indicators (16+3+4=23 dimensions in total). An M×23 dimensional original data matrix X = [x m k (m = 1 to M, k = 1 to 23), where x m k This represents the original value of the m-th fault sample on the k-th feature.

[0107] In some embodiments, outlier handling uses the 3σ principle to identify outliers and replaces them with the median; dimensionless processing uses the Z-Score method to standardize the data.

[0108] Specifically, in the outlier handling process, the "3σ principle" is first used to identify outliers, that is, for each feature k, its mean μ is calculated. k and standard deviation σ k If the feature value x of a certain sample mk Satisfy |x mk -μ k |>3σ k If the value is not found to be outlier, it is considered an outlier; secondly, for the identified outliers, the median Med of feature k is used. k Replace the outlier to prevent the mean from being affected by it, resulting in the processed matrix: X_clean = [x' mk (M×23).

[0109] Specifically, the Z-Score method is used to make the numerical values ​​of the processed matrix dimensionless.

[0110]

[0111] Where, X′ mk Let μ be the original value of the m-th fault category on the k-th index after noise reduction processing. k It is the mean of index k corresponding to all fault categories, σ k This represents the standard deviation of index k for all fault categories. After standardizing each index for each fault type, the output is the standardization matrix X_std = [x] mk ](M×23), where x” mk The value range is usually [-3, 3].

[0112] S230: Through covariance analysis and eigenvalue calculation, core features representing the preset amount of information are selected;

[0113] Specifically, preliminary dimensionality reduction will be performed, and the covariance matrix of the standardized matrix X_std output in the previous step will be calculated as follows:

[0114]

[0115] Where Cij represents the covariance between the eigenvalue of the i-th indicator and the eigenvalue of the j-th indicator.

[0116] The eigenvectors will be solved below. The eigenvalues ​​of the covariance matrix C, λ1≥λ2≥...≥λ, will be calculated using SPSS software. 25 and the corresponding eigenvectors ξ1,ξ2,...,ξ 25 (Each feature vector is 23-dimensional).

[0117] Specifically, core features will be selected. SPSS software will be used to generate scree maps, and the top n feature values ​​that reflect over 90% of all indicator information will be chosen. These n feature values ​​will be used as core indicators, and the fault categories will be classified according to their respective characteristics. Each of these n categories will be named, and the corresponding feature vectors will form a projection matrix P(2^3 × n). After the projection matrix is ​​generated, X_pca = X_std × P(M × n) can be calculated, representing the projection value of each fault sample on the n PCA core features, providing input information for the construction of the neural network model.

[0118] S240: Utilize a neural network model to learn the correlation between core features and the degree of fault damage, and calculate the importance weights of core indicators.

[0119] First, a feedforward neural network model is established. The model input information includes an input matrix and input labels. The input matrix is ​​X_FNN, whose value comes from X_pca calculated in the previous steps, and its dimensions are also kept consistent. The input model label Y_true represents the quantification value of the degree of fault damage (M×1), which is mainly based on the global fault weight W obtained in step S101. m calculate:

[0120] Y_true m =W m / W_max;

[0121] In the formula, W m Let W_m be the global fault weight corresponding to the m-th sample, and W_max be the maximum value of all global fault weights. Map the calculation results to the [0,1] interval to ensure compatibility with the output layer's Sigmoid function.

[0122] The model output information Ypred represents the predicted fault damage quantification value (M×1), ranging from [0,1], and is compared with Ytrue to calculate the loss.

[0123] The next step is to train the model. Specifically, the input matrix and input labels (X_pca, Y_true) are first divided into a training set (X_train, Y_train) and a validation set (X_val, Y_val) in an 8:2 ratio to ensure that the training set covers all fault types. At the same time, the parameters are initialized, such as W1~N(0,0.01), b1=0(1×h).

[0124] After parameter initialization, forward propagation will be performed, dividing the layer below the input layer into a hidden layer and an output layer. The hidden layer function is as follows:

[0125] H=RELU(X_FNN×W1+b1)(M×h).

[0126] The ReLU function here is a piecewise linear function. When the input value X≥0, the function output equals the input value itself; when X<0, the function output is 0. This function is intended to inject non-linear elements into the neural network, enabling the hidden layers to learn this complex correlation and overcoming the limitations of linear methods. X_FNN, W1, and b1 are all input information. The ReLU function truncates negative feature responses to 0, retaining only the feature information that positively contributes to the degree of fault damage, reducing redundant interference.

[0127] The output layer prediction function is:

[0128] Ypred=Sigmoid(H×W2+b2)(M×1).

[0129] The Sigmoid function here is a continuous, smooth sigmoid activation function. The input value can take any value, and the output value will be strictly mapped to the interval [0, 1]. The purpose of this function is to standardize the model output to a quantized value compatible with the preceding steps in the patent, providing a suitable input for calculating the loss exponent in subsequent steps.

[0130] Therefore, the formula for calculating the loss function is:

[0131]

[0132] Among them, W1 and W2 are norm values, which are mainly used to penalize excessively large weights and avoid overfitting.

[0133] Specifically, the next step will be backpropagation, starting from the error output by the model, tracing back and calculating the contribution of each parameter to the error, and then adjusting the parameters through gradient descent to reduce the error, so that the predicted value Ypred (the quantification value of the degree of failure damage) of the neural network is as close as possible to the true label Ytrue. Ultimately, the model can accurately learn the nonlinear relationship between "PCA core indicators → degree of failure damage", providing reliable model parameter support for the subsequent calculation of the importance of core indicators.

[0134] Calculate the partial derivatives of the loss function with respect to W1, b1, W2, and b2 respectively, as follows:

[0135]

[0136] Among them, H T H is the transpose of the output of the hidden layer function, ⊙ represents the element-wise product, and H is the transpose of the output of the hidden layer function. m This is the value output by the hidden layer for the m-th sample.

[0137] In subsequent parameter updates, a fixed learning rate ηlearn is set, and gradient descent is used for iteration. The upper limit of the number of iterations is specified as 50. That is, if the validation set loss Loss_val does not decrease for 50 consecutive rounds, training stops, and the optimal parameters W1*, b1*, W2*, and b2* are saved. The specific formula for iteration is shown in the following equation:

[0138]

[0139] After the iteration is completed, the iteration results are verified, requiring the mean square error of the verification set MSE_val ≤ 0.005 and the coefficient of determination R0. 2 _val≥0.92 to ensure the reliability of model fitting.

[0140] Based on the further improvements to the above method, and using the optimal weights of the neural network, the "importance weights" of the PCA core indicators are calculated. This further refines the core indicators selected in the original PCA screening step, ensuring that the final output core indicators not only meet the requirement of "information content ≥ 90%" but also possess the characteristic of "high contribution to the degree of fault damage," thus achieving the goal of extracting core indicators. Based on the optimal weight parameter W1* output from the above steps, the contribution Ik of each PCA feature k (k = 1 to n) to the output is first calculated. The calculation formula is as follows:

[0141]

[0142] in, The table represents the absolute values ​​of the weights from the k-th feature in the input layer to the j-th neuron in the hidden layer, and

[0143] S250: Based on the importance threshold, the core indicators are screened a second time, and the core evaluation indicator set and corresponding weights are output.

[0144] Based on further improvements to the above method, a second screening of the core indicators is performed. An importance threshold β = 0.05 is set, ensuring that Ik ≥ β. If fewer than 5 Ik values ​​meet this condition, at least the top 5 largest Ik values ​​are retained to guarantee the representativeness of the core indicators and avoid insufficient data collection due to too few dimensions. After calculation, the final core indicator set F_core = [f1, f2, ..., ft] (t ≤ n) is output, along with the corresponding importance weight vector I_core = [I1, I2, ..., It].

[0145] Based on the further improvement of the above method, the feature vector F_core corresponding to the PCA feature index is combined with the weight vector I_core for weighting to obtain the "core index comprehensive score" S for each fault sample:

[0146]

[0147] Sm represents the comprehensive score of the core indicators for the m-th sample.

[0148] S300: Based on disaster weight, failure mode weight and importance weight of core indicators, a loss index calculation model is established to calculate the basic loss index for failure sample data and preset the loss level range.

[0149] In some embodiments, the specific steps for calculating the basic loss index include:

[0150] S301: Based on power grid maintenance guidelines and historical statistical characteristics, preset the boundary conditions for loss levels;

[0151] In some embodiments, the loss level boundary conditions are dynamically calculated based on the index score quantiles, including: index ranges for minor loss, moderate loss, and severe loss.

[0152] Based on the core features obtained in the previous step S200, the boundary conditions for various loss degree discriminations are calculated:

[0153] Level I: Minor Loss Boundary The lower limit is 0.

[0154] Level II: General Loss Boundary

[0155] Level III: Severe Loss Boundary The maximum is 100.

[0156] Wherein, S5% and S95% are the quantile boundaries of the "core indicator comprehensive score" S calculated in the previous step S200. First, a set of loss severity grading intervals is artificially defined as shown in Table 5 below:

[0157] Table 5. Loss Severity Classification

[0158] Loss Level Qualitative description Loss Index Range Level I (Minor Damage) Some accessories are damaged, but the core functions are not affected. 0～L1 Level II (General Loss) Some core metrics are abnormal, and functionality is limited. L1~L2 Level III (Severe Damage) Core components damaged, functionality interrupted L2～100

[0159] S302: Combine the importance weights of core indicators and the weights of failure modes to calculate the indicator-level weighted score for each failure sample.

[0160] After classifying the loss levels, based on the further improvement of the above method, for each fault sample m (corresponding to the m-th type of fault in step S100), the weighted score of the indicator layer is first calculated by combining the core indicator score Smk and the indicator importance weight Ik. The calculation formula is as follows:

[0161]

[0162] S303: Calculate the fault layer weighted score based on the indicator layer weighted score and the fault concurrency coefficient.

[0163] Further calculate the weighted score for the fault layer:

[0164]

[0165] S304: Normalize the weighted scores of all fault layers, map them to a preset interval, and generate a basic loss index.

[0166] Specifically, the weighted score for the fault layer is further calculated:

[0167]

[0168] Fc is the fault concurrency coefficient in step S200, which represents the correlation of faults.

[0169] S400: Performs nonlinear calibration on the basic loss index and integrates multidimensional features to generate a comprehensive loss index.

[0170] In some embodiments, the specific steps for nonlinear calibration of the basic loss index include:

[0171] S410: Core indicators are classified into threshold indicators, linear indicators, and saturation indicators.

[0172] Among them, threshold-type indicators are those in which the loss index will increase sharply after the actual value is lower or higher than the threshold, such as breakdown voltage and insulating oil acidity level; linear indicators are those in which the loss changes linearly with the indicator value, such as winding DC resistance imbalance rate and no-load loss; saturation-type indicators are those in which the loss index growth rate slows down after the indicator value reaches a certain level, such as cooling fan start-up success rate and insulation resistance recovery rate per unit time.

[0173] S420: Design a corresponding calibration function for each core indicator type and apply the calibration function to the core indicator score.

[0174] In some embodiments, the calibration function includes: an exponential function, a hyperbolic tangent function, and a saturation function; wherein the exponential function is used for threshold-type indicators; the hyperbolic tangent function is used for linear-type indicators; and the saturation function is used for saturation-type indicators.

[0175] Threshold-type index calibration function: Where X0 is the index threshold, and k1 = Y predm .

[0176] The calibration function for the linear index is f2(x) = tanh(0.5x), which uses the hyperbolic tangent function to map the linear score to the range of [-1, 1].

[0177] Calibration function for saturation index Where k3 = 1 + I k .

[0178] Apply the corresponding calibration function to the score Smk of each core indicator and recalculate the weighted score of the indicator layer. The calibration loss index LIcal is derived, and its calculation formula is as follows:

[0179]

[0180] S430: Recalculate the weighted score of the calibrated index layer and generate the calibration loss index by combining the failure mode weights.

[0181] Further improvements to the above method will be made to calculate the basic index weight, calibration index weight, and correlation index weight, and then combine the three types of weights to calculate the final comprehensive loss index.

[0182] Among them, the basic index weights are:

[0183] Calibration index weight: δ = 0.5 × max(I) core );

[0184] Correlation index weight: η = 1 - α - δ.

[0185] S440: Integrates the basic loss index, calibration loss index, and disaster-related weights to generate a comprehensive loss index.

[0186] The final overall loss weight is: LI final =α×LI base +δ×LI cal +η×LI rel .

[0187] Among them, LI is specified final ∈[0,100], if the calculated result L i <0, then L i Calculated as 0; if L i If L > 100, then L i Calculated at 100.

[0188] S500: The comprehensive loss index is optimized by a random forest regression model, and the final loss prediction index is output. The equipment repair priority is determined based on the final loss prediction index.

[0189] In some embodiments, the specific steps for optimizing the comprehensive loss index include:

[0190] S501: Construct a multidimensional correlation matrix that includes core loss features, weight features, and device scenario features.

[0191] Specifically, a three-dimensional correlation matrix is ​​first constructed, comprising core loss features, weight and influencing factor features, and device and scene features, as follows:

[0192] The first dimension is the core loss feature, whose parameters include LIfinal, LIbase, LIcal, and LIrel.

[0193] The second dimension is the characteristics of weights and influencing factors, whose parameters include the fusion weights α, δ, and η, as well as the key influencing factors Ik and Wb.

[0194] The third dimension is equipment and scenario characteristics. Equipment attribute evaluation parameters include disaster resistance level Lr and service life Ys, while scenario characteristic evaluation parameters include fault propagation risk level Rs and fault correlation coefficient Fc.

[0195] By combining the above parameters, an M*12 dimensional feature matrix is ​​constructed.

[0196] S502: Perform outlier correction on the comprehensive loss index to ensure the rationality of data distribution.

[0197] Based on further improvements to the above method, outliers in the final calculation result of S103 are corrected a second time. The threshold for identifying outliers is set at 10%, meaning that if the LIfinal of a sample exceeds or does not reach 10% of the dynamic range Int, it is determined to be an outlier. Linear interpolation is used to correct the outliers, as shown in the following formula:

[0198] LI final,crror =α×LI base +δ×LI cal +η×clip(LI rel (L1×0.9, L3×1.1).

[0199] The Clip function is a truncation function that limits LIrel to within ±10%.

[0200] S503: It uses a random forest model and an XGBoost model for parallel training and dynamically adjusts the model parameters.

[0201] A random forest fusion XGboost model will be constructed, whose parameters will be dynamically adjusted based on the fusion weights and feature importance in step S300, and will inherit the nonlinear calibration logic of step S300 to improve the optimization accuracy.

[0202] In some embodiments, the parameters of the random forest regression model are dynamically set according to the fusion weights and disaster resistance level, including: the number of decision trees, the maximum depth, and the learning rate; the constraint for training the random forest regression model is to ensure that the predicted level matches the initial level.

[0203] First, define the parameters of the random forest model:

[0204] Let the number of decision trees be expressed as n. estimators,RF =150+50×max(α,β,γ).

[0205] That is, the larger the maximum value of the fusion weight, the more trees there are, in order to adapt to the weight differences represented by each parameter; let the maximum depth be: max depth,RF =10+round(Lr).

[0206] The higher the disaster resistance level, the less depth it represents.

[0207] Other parameters are assigned the following values: batch size = 32, minimum leaf node sample count = 5, regularization coefficient = 0.0001.

[0208] In the XGBoost model, the learning rate η XGB =0.005×(1+γ);

[0209] Number of trees n estimators,XGB =200 + 100 × δ;

[0210] Other parameters are as follows: maximum depth = 8, subsample ratio = 0.8, column sampling ratio = 0.7, L2 regularization coefficient = 0.001.

[0211] S504: Calculate the model fusion weights based on validation set performance and weight allocation.

[0212] Based on the above method, a dual-model parallel training is adopted, with the training set and validation set divided in an 8:2 ratio. The training set covers all loss levels of S103 to ensure the generalization ability of the model. The goal is to minimize the mean squared error (MSE) of the validation set, while setting the constraint that the model output is consistent with the S103 level (i.e., the matching rate between the predicted level and Gradefinal is ≥90%).

[0213] Based on the validation set performance and the fusion weights calculated in S103, the weights of the two models are calculated:

[0214]

[0215] Where, ω RF +ω XGB =1.

[0216] S505: Outputs the final loss prediction index through weighted average.

[0217] The expression for the loss prediction exponent after fusing the two models is as follows:

[0218] LI pred,fuse =ω RF ×LI pred,RF +ω XGB ×LI pred,XGB .

[0219] As can be seen from the above embodiments, this application provides a method for assessing the degree of damage to main transformer equipment after a disaster, including: encoding different natural disasters that may cause damage to main transformer equipment according to categories, constructing a system of factors affecting main transformer equipment failure under multiple natural disaster types, and analyzing the weight of each factor; listing a series of indicators for assessing the post-disaster repair status of equipment based on historical disaster damage data and equipment operating status, and performing dimensionless processing, and using principal component analysis to analyze and identify its core indicators; establishing a calculation model for the degree of equipment loss, integrating disaster weights and core indicators, and calculating a linearized loss index; finally, through a random forest regression model, introducing a training sample set, clarifying the fusion calculation weights and improving the accuracy of index calculation, and determining the final loss index.

[0220] This application not only transforms the qualitative experience of domain experts into quantitative weights through the analytic hierarchy process (AHP), clarifying the destructive contribution of different disaster-induced failure modes and effectively solving the problems of strong subjectivity and reliance on personal experience in traditional assessments, but also utilizes principal component analysis (PCA) to reduce the dimensionality of massive, multidimensional fault monitoring indicators, extracting the most representative core assessment dimensions and overcoming the shortcomings of redundant indicators and vague assessment focus in traditional methods. Furthermore, by introducing a random forest regression model to integrate disaster weights and core indicators, it can effectively characterize the complex nonlinear relationship between disaster, failure, and main transformer loss, significantly improving the accuracy of loss index calculation and model adaptability. This comprehensive assessment method ultimately outputs a unified and quantifiable loss index and level, providing a clear and reliable decision-making basis for post-disaster repair priority determination and dynamic scheduling of human and material resources, thereby achieving optimal allocation of maintenance resources and improving the overall efficiency and resilience of power system post-disaster recovery.

[0221] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein.

Claims

1. A method for quantitatively evaluating the loss degree of a post-disaster main transformer equipment, characterized in that, The method comprises the following steps: Obtain natural disaster category data and main transformer equipment failure data, construct a disaster-caused failure influence factor system, calculate disaster weights and failure mode weights; Based on historical disaster damage data and equipment operating status, list equipment maintenance evaluation indicators, preprocess the data of the equipment maintenance evaluation indicators, identify a core evaluation indicator set, and calculate the importance weights of the core indicators; According to the disaster weights, the failure mode weights, and the importance weights of the core indicators, establish a loss index calculation model, calculate the basic loss index of the failure sample data, and preset the loss level interval; Nonlinearly calibrate the basic loss index, and generate a comprehensive loss index by fusing multi-dimensional features; Optimize the comprehensive loss index through a random forest regression model, output the final loss prediction index, and determine the equipment repair priority according to the final loss prediction index.

2. The method of claim 1, wherein, The specific steps of constructing the disaster-caused failure influence factor system comprise: Encode the natural disaster categories to form a disaster category layer; For each natural disaster category, match the corresponding failure mode to form a failure mode layer; Based on expert scoring data, construct a disaster damage degree comparison matrix and a failure damage probability matrix; Calculate the weight of each natural disaster category through geometric mean and normalization processing; Calculate the weight of each failure mode under each natural disaster category and perform consistency test; Output a global failure weight ranking, which is used to identify the equipment repair priority.

3. The method of claim 1, wherein the method further comprises: The specific steps of identifying the core evaluation indicator set comprise: Integrate equipment maintenance major indicators, failure correlation indicators, and equipment basic property indicators to form an original indicator matrix; Perform outlier processing and dimensionless processing on the original indicator matrix to generate standardized data; Through covariance analysis and eigenvalue calculation, filter the core features representing the preset information amount; Use a neural network model to learn the association between the core features and the failure damage degree, and calculate the importance weights of the core indicators; Based on the importance threshold, perform secondary screening on the core indicators to output the core evaluation indicator set and the corresponding weights.

4. The method of claim 1, wherein the method further comprises: The specific steps of calculating the basic loss index comprise: According to the grid maintenance guide and historical statistical characteristics, preset the loss level boundary conditions; Combine the importance weights of the core indicators and the failure mode weights to calculate the indicator layer weighted score of each failure sample; Based on the indicator layer weighted score and the failure concurrency coefficient, calculate the failure layer weighted score; Normalize all the failure layer weighted scores, map them to the preset interval, and generate the basic loss index.

5. The method of claim 1, wherein the method further comprises: The specific steps of nonlinearly calibrating the basic loss index comprise: Divide the core indicators into threshold type indicators, linear type indicators, and saturated type indicators according to types; Design a corresponding calibration function for each core indicator type, and apply the calibration function to the core indicator score; Recalculate the calibrated indicator layer weighted score, and generate a calibrated loss index by combining the failure mode weights; Fuse the basic loss index, the calibrated loss index, and the disaster correlation weight to generate the comprehensive loss index.

6. The method of claim 1, wherein the method further comprises: The specific steps of optimizing the comprehensive loss index include: Constructing a multi-dimensional correlation matrix containing core loss features, weight features, and device scene features; Performing outlier correction on the comprehensive loss index to ensure reasonable data distribution; Parallel training of random forest model and XGBoost model with dynamic adjustment of model parameters; Calculating model fusion weight based on validation set performance and weight allocation; Outputting the final loss prediction index through weighted average.

7. The method of claim 3, wherein the method further comprises: The outlier processing uses the 3σ principle to identify outliers and replaces them with the median; The dimensionless processing uses the Z-Score method to realize data standardization.

8. The method of claim 4, wherein the method further comprises: The loss level boundary condition is dynamically calculated based on the index score quantile, including the index range of slight loss, general loss, and severe loss.

9. The method of claim 5, wherein the method further comprises: The calibration function includes exponential function, hyperbolic tangent function, and saturation function; The exponential function is used for the threshold type index; The hyperbolic tangent function is used for the linear type index; The saturation function is used for the saturation type index.

10. The method of claim 6, wherein the method further comprises: The parameters of the random forest regression model are dynamically set according to the fusion weight and disaster resistance level, including the number of decision trees, maximum depth, and learning rate; The constraint condition for training the random forest regression model is to ensure that the predicted level matches the initial level.

Images