Power failure prediction method based on big data model training reasoning with explainable important features

Abstract

The invention discloses an important feature interpretable big data model training reasoning-based power failure prediction method. The method comprises the following steps of firstly, collecting a power data set; training an original prediction model by using the training set; calculating the contribution degree of the feature of each data sample in the training set to a prediction result by adopting a feature importance evaluation algorithm to obtain a global importance score of each data sample; based on the global importance score, selecting an important feature subset through a feature screening method; training a proxy model with inherent interpretability by using the important feature subset; inputting a to-be-predicted sample into the original prediction model and the proxy model in parallel; and finally, outputting a final prediction value by taking a prediction result of the original prediction model as a reference, and meanwhile, outputting interpretable data generated by the agent model, so that the complex big data model can efficiently and stably predict faults in the power system on the premise of basically not losing prediction precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of big data and power systems, and in particular to a power fault prediction method based on important feature interpretable big data model training and inference. BACKGROUND

[0002] In the process of power system operation, equipment fault prediction is the core link to ensure the safety and stability of the power grid and reduce operation and maintenance costs. With the deep integration of big data technology and artificial intelligence, the power fault prediction scheme based on complex big data models (such as deep learning and integrated learning models) has achieved high prediction accuracy in transformer fault early warning, transmission line icing prediction, and distribution equipment anomaly identification, etc. scenarios, with its strong fitting ability for massive multi-source power data (such as power generation side load, transmission line parameters, equipment operating status, environmental factors, etc.), providing technical support for the transformation of power operation and maintenance from "after-the-fact maintenance" to "pre-emptive prevention".

[0003] However, in the actual power operation and maintenance scenario, the existing fault prediction scheme based on complex big data models faces key technical bottlenecks: first, the "black box" nature of complex models leads to a lack of interpretable basis for prediction results. The operation and maintenance team needs to identify "associated factors of increased fault risk" (such as transformer winding temperature anomalies, insulation oil degradation, etc.) to develop maintenance plans, but existing models only output results and cannot present associated factors and contribution, making it difficult for the operation and maintenance team to accurately locate the fault location; second, the existing scheme for improving interpretability is not adaptable: inherent interpretable models (such as linear regression and shallow decision trees) have weak fitting ability for the nonlinear and time series characteristics of power systems, with prediction accuracy lower than industrial requirements (such as transformer fault prediction accuracy often less than 80%), making it difficult to identify early hidden faults; post-explanation methods based on LIME and SHAP also have high computational cost (single inference takes more than 0.5 seconds, which does not meet real-time requirements), poor explanation stability (difference rate of explanations for the same sample exceeds 15%), and incomplete logic chain (only reflects local feature contribution), resulting in false positives and false negatives in reality, decision-making errors, and other phenomena, which greatly wastes operation and maintenance resources.

[0004] In view of the deficiencies of the existing technology, there is an urgent need for a fault prediction method that adapts to the actual needs of the power system, provides stable and complete interpretable basis for the prediction results under the premise of ensuring prediction accuracy and inference efficiency, solves the technical contradiction of "unreliable black box model" and "low accuracy of interpretable model", and supports the scientificity and reliability of power operation and decision-making. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art, and provide an important feature interpretable big data model training inference based power failure prediction method, which can realize efficient and stable prediction of power system failure under the premise of basically not losing prediction accuracy.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0007] An important feature interpretable big data model training inference based power failure prediction method,

[0008] Through important feature screening and proxy model training in the offline stage, combined with double-channel inference in the online stage, the synchronous output of prediction results and decision basis is realized, and the specific steps are as follows:

[0009] I. Training stage (offline)

[0010] (1) Large-scale training data acquisition and preprocessing. First, in the training stage, the present method collects large-scale data from the original data. The data sources can be power generation side data, power transmission side data, power distribution side data and power equipment full life cycle data in the field of power engineering. Then, the batch processing of abnormal values and missing values of the collected large-scale data is carried out. In order to improve the data quality, the present method identifies and eliminates abnormal data by IQR method; for the abnormal values and missing values in the data, further combined with its label attribute (continuous variable / discrete variable), the mean or mode is used for standardized filling. Then, the data is processed by feature encoding and normalization. The present method respectively adopts independent encoding and label encoding for unordered category and ordered category features, and extracts time stamp features for time series features. In addition, the continuous features are standardized by Z-Score or normalized by Min-Max. Finally, after the arrangement, the qualified data set can be obtained, which is used as the training set of the subsequent model by the present method.

[0011] (2) Training a high-performance original prediction model. This method uses the complete training data set pre-processed in step (1) to train an original prediction model with different application scenario characteristics, with prediction accuracy as the core target. First, task classification is determined. This method distinguishes between classification tasks and regression tasks by whether the prediction label is discrete and whether the number of prediction label categories is greater than 20. Next, candidate prediction model selection and training are performed. This method selects XGBoost, LSTM, Transformer, CNN, and Logistic as candidate models. These five models can cover more than 95% of big data prediction scenarios and can flexibly respond to complex prediction needs in different scenarios and fields. This method randomly divides the training set and validation set in an 8:2 ratio for all candidate models, and uses Bayesian optimization to search the hyperparameter space, with each model optimized for ≥50 iterations to ensure that the hyperparameters reach local optimality. Then, record the result indicators of each model. If the task is a classification task, this method calculates the precision, recall, and F1 score for each candidate model, as follows:

[0012]

[0013] where T p is the true positive, representing the number of times the model predicts a true positive as positive; F N is the false negative, representing the number of times the model predicts a true positive as negative; and F P is the false positive, representing the number of times the model predicts a true negative as positive.

[0014] If the task is a regression task, this method calculates the root mean square error, mean absolute error, and coefficient of determination for each candidate model, as follows:

[0015]

[0016] where y i is the true value of the i-th sample, is the predicted value of the i-th sample by the model, is the average value of all sample true values, and n is the number of samples.

[0017] Finally, the model with the best prediction accuracy is selected as the original prediction model of this method, serving as the "accuracy main channel" in the subsequent inference stage, providing high-accuracy prediction results.

[0018] (3) Calculate feature importance. The feature importance evaluation algorithm uses the SHAP algorithm.

[0019] Using the complete training dataset preprocessed in step (1) as input, calculate the marginal contribution (SHAP) value of all features to the prediction result of this sample, and for each feature, take the mean absolute error (MA-SHAP) of its SHAP values ​​across all samples as the global importance score of that feature, as follows:

[0020]

[0021] Where f is the target feature, n is the total number of training samples, and |SHAP i,f | represents the SHAP value of feature f in the i-th sample.

[0022] (4) Selecting an important feature subset. The important feature selection methods used in this method include Top-K selection, threshold selection, and elbow rule. There are no hard and fast requirements for the selection, as long as the performance difference is within an acceptable range in the final verification stage; otherwise, the parameters should be modified or the selection method should be changed.

[0023] The elbow rule is used to determine the minimum number of features with a cumulative contribution of ≥90%, ensuring that the selected feature subset covers the core information predicted by the model. This method sorts all features according to their global importance score from highest to lowest, and calculates the cumulative contribution of the top m features sequentially, using the following formula:

[0024]

[0025] Where k is the number of features currently selected, M is the total number of features in the original model, and f i Let i be the i-th feature after sorting, and CumContrib(k) be the cumulative contribution of the first k features.

[0026] This method plots a line graph with the number of features k on the horizontal axis and the cumulative contribution CumContrib(k) on the vertical axis to observe the trend of the curve. By manually selecting the smallest k value at the "elbow inflection point and with a cumulative contribution ≥ 90%" as the number of features in the important feature subset, it ensures that no core prediction information is missed.

[0027] (5) Training an interpretable surrogate model. This method uses the "important feature subset" selected in step (3) and its corresponding sample labels to train a surrogate model with a simple structure and inherent interpretability. First, candidate surrogate models are selected and trained. Logistic regression model, linear regression model, shallow decision tree model with a depth of ≤3 layers and generalized additive model are selected as candidate models for this method, and the optimal model is selected. The specific steps are similar to those in step (2).

[0028] The method also needs to verify the performance of the proxy model. The performance of the proxy model is verified using a test set consistent with the original model. If it is a classification task, it needs to ensure that the accuracy difference with the original model is less than or equal to 3%, and if it is a regression task, it needs to ensure that the mean absolute error difference with the original model is less than or equal to 5%. If the consistency is not up to standard, return to step (4) to adjust the important feature screening strategy. Finally, in order to ensure that the proxy model has the ability to explain, the linear model retains the coefficients of each feature, and the decision tree records the split features and thresholds of each node to directly output the explainability data. The proxy model serves as an "explanation channel" for the subsequent reasoning stage, providing intuitive decision-making basis for the prediction results of the original model.

[0029] II. Reasoning stage (online)

[0030] (1) Preprocessing of the sample to be predicted. When a new sample to be predicted is obtained, first perform the same data preprocessing operations as in the training stage, including data cleaning, missing value filling, feature encoding, and standardization, etc. processes to ensure that the input data format is highly matched with the training stage.

[0031] (2) Double-channel parallel reasoning. First, determine whether the current task is a classification task or a regression task according to whether the prediction label is discrete and whether the number of discrete labels exceeds 20. Then, the preprocessed sample is input into two reasoning channels in parallel:

[0032] Channel one (precision main channel): input the complete feature sample after preprocessing into the trained original prediction model to obtain a high-precision first prediction result, which serves as the benchmark for the final prediction value.

[0033] Channel two (explanation channel): extract the corresponding k important features from the preprocessed sample and input them into the trained proxy model. The proxy model will output a second prediction result, which needs to have a high degree of consistency with the first prediction result. On the other hand, it will naturally generate explainability data. For linear models, the explainability data is represented by the coefficients and weights of each feature. For decision tree models, the explainability data is the specific decision path, etc.

[0034] Result synthesis and output. The system finally outputs the final prediction value based on the prediction result of the original prediction model, and outputs the explainability data generated by the proxy model, including the specific contribution value, contribution direction, and visualization chart of each important feature. Based on the explainability data, the prediction result can be reasonably explained and explained.

[0035] The present application has the following characteristics and beneficial effects:

[0036] The application realizes efficient interpretability of complex big data model prediction results by constructing a double model collaborative architecture of "high-precision original prediction model + interpretable proxy model". On the one hand, the architecture retains the high-precision characteristics of the complex original prediction model, and at the same time, the lightweight proxy model training is completed in the offline stage, which significantly reduces the computational overhead of explanation generation in the online inference stage, ensuring the real-time nature of the explanation process. On the other hand, the proxy model can directly output explanation information with clear physical meaning as an intuitive and understandable decision basis. Moreover, the application has good model compatibility and does not depend on a specific model type, so it can be flexibly adapted to various combinations of black box or white box models. By applying this method to the power operation and maintenance scene, the operation and maintenance team can achieve high-accuracy power system fault prediction and at the same time accurately locate the fault position for recovery. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The method flowchart of the embodiment of the application is shown.

[0038] Figure 2 The double-channel model framework method diagram in the embodiment of the application is shown.

[0039] Figure 3 The "feature number k-cumulative contribution degree" line graph drawn by the elbow rule in the embodiment of the application is shown. DETAILED DESCRIPTION

[0040] The application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0041] A power fault prediction method based on important feature interpretable big data model training and inference, as shown in Figure 1 includes the following steps:

[0042] Step 1, collect power generation side, power transmission side, power distribution side and power equipment full life cycle data in the power engineering field and preprocess to obtain a power dataset, and divide the power dataset into a training set and a validation set.

[0043] Specifically, in the power transformer equipment fault prediction scenario in the embodiment, the data is derived from 150,000 transformer operation state record data sets collected in a regional power grid dispatching system, each record contains device basic information (model specifications, operation life, etc. 10 characteristics), real-time operation data (winding temperature, load current, etc. 12 characteristics), environmental parameters (environmental temperature, humidity, etc. 4 characteristics) and "whether to occur fault" label. This method first uses the IQR method to remove 4800 abnormal samples of "winding temperature > 150℃" and "load current > 150% of rated value", and processes the missing values of the remaining 145200 samples according to the feature type. The continuous type feature is filled with mean value, and the discrete type feature is filled with mode. Then, the unordered category features (such as the manufacturer) are encoded with one-hot encoding, the ordered category features (such as the insulation grade) are encoded with label encoding, and the continuous type features (such as the load current) are compressed to the [0, 1] interval using Min-Max normalization. Finally, a data set containing 26 input features and 1 label is formed, and is divided into a training set (116160) and a test set (29040) according to 8:2.

[0044] Step 2, use the preprocessed training set to train multiple candidate prediction models, and select the best model as the original prediction model through performance evaluation.

[0045] Specifically, in the embodiment, first, according to the "whether to occur fault" label, it is determined that it is a binary classification task because it is discrete and the number of categories is 2 (≤20). Then, five candidate models of XGBoost, LSTM, Transformer, CNN and Logistic regression are selected. For each model, the hyperparameters are searched using Bayesian optimization (iterating 50 times, such as the optimal parameters of XGBoost are learning rate 0.07, tree depth 5, and estimators = 250). The training set and the validation set are divided according to 8:2 to calculate the precision, recall and F1 score of each model.

[0046] Specifically, if the task is a classification task, the method calculates the corresponding precision, recall and F1 score for each candidate model, and the calculation formula is as follows:

[0047]

[0048] Where, T p is the true positive, which represents the number of times that the model predicts the true positive as positive; F N is the false negative, which represents the number of times that the model predicts the true positive as negative; F P is the false positive, which represents the number of times that the model predicts the true negative as positive.

[0049] If the task is a regression task, the method calculates the root mean square error, mean absolute error and coefficient of determination for each candidate model, and the calculation formula is as follows:

[0050]

[0051] where y i is the true value of the i-th sample, is the predicted value of the i-th sample by the model, is the average value of all sample true values, and n is the number of samples.

[0052] Finally, the model with the best prediction accuracy is selected as the original prediction model of the method by comparing the index results, which provides a high-credibility prediction result as the "accuracy main channel" in the subsequent reasoning stage. The results are shown in Table 2:

[0053] Table 2. Comparison of prediction performance of each candidate model

[0054]

[0055]

[0056] According to the results, the XGBoost model has the best performance (precision 0.892, recall 0.875, F1 score 0.883), so it is determined as the original prediction model of the "accuracy main channel", and its final performance on the test set is precision 0.887, recall 0.869, and F1 score 0.878.

[0057] It should be noted that the five models selected in this embodiment can cover more than 95% of big data prediction scenarios and can flexibly meet the complex prediction needs of different scenarios and fields.

[0058] Step 3. Calculate the contribution of each input feature in the training set to the prediction result using a feature importance evaluation algorithm to obtain the global importance score of each data feature.

[0059] Specifically, in this embodiment, the SHAP algorithm is used, and the method takes the complete training data set after preprocessing in step (1) as input to calculate the marginal contribution SHAP value of all features to the prediction result of the sample. For each feature, take the average absolute error (MA-SHAP) of the SHAP value of all samples as the global importance score of the feature, and the formula is:

[0060]

[0061] where f is the target feature, n is the total number of training samples, and |SHAP i,f is the SHAP value of feature f in the i-th sample.

[0062] It can be understood that the feature importance is calculated and the important feature subset is screened. According to the correspondence between the "prediction model-feature importance algorithm", the XGBoost model matches the SHAP / LIME algorithm, and the SHAP algorithm is selected in this embodiment. The SHAP value of each feature on all samples is calculated by taking the preprocessed complete training set as input, and the global importance score is determined by the average absolute SHAP value (MA-SHAP). The top 3 important features are "winding temperature" (MA-SHAP = 0.312), "load current" (MA-SHAP = 0.278), and "insulating oil breakdown voltage" (MA-SHAP = 0.245).

[0063] It should be noted that the global importance score calculation method can also use LIME, IG (Integrated Gradients) and PFI (Permutation Feature Importance) algorithms. The specific selection can be made according to the specific scene. In this embodiment, the different prediction models of step (2) and the different feature importance evaluation algorithms of this step have a corresponding relationship in the application scene. For example, the Logistic regression model and the PFI algorithm are most suitable for simple linear relationship scenes. Therefore, if the prediction model selects the Logistic regression model, the corresponding PFI algorithm can also achieve good evaluation effect. The corresponding relationship is shown in Table 1:

[0064] Table 1. Association table of prediction model and feature importance evaluation algorithm

[0065]

[0066] The method randomly selects one of the optimal feature importance evaluation algorithms corresponding to the final determined prediction model in the same scene to calculate the contribution of each feature to the prediction result.

[0067] Step 4, based on the global importance score, an important feature subset is selected by a feature screening method.

[0068] The important feature screening method used in this method includes Top-K selection, threshold screening and elbow rule. It should be noted that there is no hard requirement for the selection of the screening method. As long as the performance difference meets the acceptable range in the final verification link, otherwise the parameters are modified or the screening method is replaced.

[0069] In this embodiment, the elbow rule is used to determine the minimum number of features with a cumulative contribution degree of ≥90%, so as to ensure that the screened feature subset can cover the core information of the model prediction. The method sorts all features in descending order of global importance score, and calculates the cumulative contribution degree of the first m features in turn. The formula is:

[0070]

[0071] Where k is the number of features currently selected, M is the total number of features in the original model, and f i Let i be the i-th feature after sorting, and CumContrib(k) be the cumulative contribution of the first k features.

[0072] This method plots a line graph with the number of features k on the horizontal axis and the cumulative contribution CumContrib(k) on the vertical axis to observe the trend of the curve. By manually selecting the smallest k value at the "elbow inflection point and with a cumulative contribution ≥ 90%" as the number of features in the important feature subset, it ensures that no core prediction information is missed.

[0073] In this embodiment, as Figure 3 The elbow method is used for screening. First, the 26 features are sorted from highest to lowest MA-SHAP score. Then, the cumulative contribution of the top k features is calculated using the formula, and a line graph of "feature count k - cumulative contribution" is plotted, as shown below. Figure 2 As shown. Observation revealed that when k=8, the cumulative contribution reached 93.85% (≥90%) and the subsequent growth slowed down. Therefore, the important feature subset was determined to include the first 8 features (including winding temperature, load current, etc.) to ensure that the core information for fault prediction can be covered.

[0074] Step 5: Train an agent model with inherent interpretability using the aforementioned important feature subset, and then verify its performance.

[0075] Specifically, a simple surrogate model with inherent interpretability is trained using the selected "important feature subset" and its corresponding sample labels. First, candidate surrogate models are selected and trained. Logistic regression, linear regression, shallow decision tree models with a depth of ≤3 layers, and generalized additive models are selected as candidate models for this method, and the optimal model is selected. The specific steps are similar to step (2).

[0076] This method also requires performance validation of the surrogate model. The performance of the surrogate model is validated using a test set consistent with the original model. For classification tasks, the accuracy difference from the original model must be ≤3%; for regression tasks, the mean absolute error difference must be ≤5%. If consistency is not met, return to step (4) to adjust the important feature selection strategy. Finally, to ensure the surrogate model has interpretability, the linear model retains the coefficients of each feature, and the decision tree records the splitting features and thresholds of each node to directly output interpretable data. This surrogate model serves as the "interpretation channel" for subsequent inference stages, providing an intuitive decision-making basis for the prediction results of the original model.

[0077] In this embodiment, four candidate surrogate models were selected: logistic regression, shallow decision tree with a depth of ≤3 layers, linear regression, and generalized additive model. The models were trained using an important feature subset (9 features) and corresponding labels. The performance was verified using a test set, requiring the accuracy difference between the surrogate model and the original XGBoost model to be ≤3%. The results showed that both the logistic regression model (accuracy 0.887, difference 2.1%) and the shallow decision tree (accuracy 0.882, difference 2.6%) met the standard. Considering that logistic regression is more intuitive in its interpretation (feature coefficients can be directly output), it was determined as the surrogate model for the "interpretation channel", and its feature coefficients (such as winding temperature coefficient 0.963, load current coefficient 0.875, etc.) were retained as the basis for interpretation.

[0078] Step 6: Obtain the sample to be predicted and preprocess it according to the method in Step 1 to obtain the preprocessed sample to be predicted.

[0079] Step 7: The preprocessed samples to be predicted are input in parallel into the original prediction model and the surrogate model; the original prediction model outputs the first prediction result, and the surrogate model outputs the second prediction result and interpretability data.

[0080] Specifically, real-time monitoring data from one 110kV transformer was selected (model S11-50000 / 110, 8 years in operation, current winding temperature 85℃, etc.). After inspection, no anomalies or missing data were found. Subsequently, this method confirmed that the task was binary classification, and then the preprocessed samples were input in parallel into two channels, such as... Figure 2 As shown:

[0081] Channel 1 (precision channel) inputs the complete 26-dimensional feature vector into the original XGBoost model and outputs a failure probability of 0.86 (≥0.5 threshold), predicting the result as "failure may occur".

[0082] Channel 2 (Explanation Channel) inputs the 9-dimensional important feature vector into the logistic regression surrogate model and outputs a fault probability of 0.83 (3.49% difference from the result of Channel 1, showing high consistency). At the same time, it generates an explanation based on the feature coefficients: "The winding temperature is 85℃ (coefficient 0.963), the load current reaches 92% of the rated value (coefficient 0.875), and the insulation oil breakdown voltage is 32kV (coefficient -0.742), which is lower than the standard value, resulting in a high risk of fault."

[0083] Step 8: Output the final predicted value based on the prediction results of the original prediction model, and simultaneously output interpretable data generated by the surrogate model. This interpretable data includes the specific contribution value, contribution direction, and visualization charts for each important feature. Based on this interpretable data, the prediction results can be reasonably explained.

[0084] Finally, the system synthesizes the output, arranges the final conclusion as "110kV S11-50000 / 110 type transformer fault risk level is 'high', and it is recommended to arrange special maintenance within 48 hours" based on the original model result, and attaches the explanation generated by the proxy model "the current winding temperature of the transformer is 85℃ (feature contribution coefficient 0.963, higher than the safety threshold 75℃, increasing the fault risk by 42%), the load current reaches 92% of the rated value (contribution coefficient 0.875, long-term high load accelerates equipment aging, increasing the fault risk by 35%), and the insulation oil breakdown voltage is 32kV (contribution coefficient-0.742, lower than the standard value 40kV, weakening the insulation capacity, indirectly increasing the fault risk by 28%)", meeting the requirement of power safety supervision for decision explainability. Based on the result, the operation and maintenance team can maintain the 110kV S11-50000 / 110 type transformer, and specially check and repair the possible abnormalities of the current winding temperature, load current and insulation oil breakdown voltage.

[0085] To further verify the implementation effect of the method in the field of electric power engineering, batch reasoning is performed on 29040 test set samples, and the results show that the original model F1 score is 0.899 (meeting the high precision requirement), the proxy model and the original model prediction consistency rate is 95.6% (high explain consistency), the online reasoning average time is 0.02 seconds / sample (better than the traditional post-explanation method, meeting the real-time requirement), the same sample repeated reasoning 100 times is completely consistent (good stability), and the explanation generated by the method can be directly provided to the operation and maintenance personnel and the supervision agency, solving the compliance and practicality pain points of electric power equipment fault prediction, and fully verifying the effectiveness of the method in the field of electric power engineering.

[0086] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A power fault prediction method based on training and inference using a big data model with interpretable key features, characterized in that, Includes the following steps: Step 1: Collect and preprocess data from the power generation, transmission, and distribution sides, as well as the entire lifecycle of power equipment in the field of power engineering, to obtain a power dataset. Divide the power dataset into a training set and a validation set. Step 2: Train multiple candidate prediction models using the preprocessed training set, and select the best model as the original prediction model through performance evaluation. Step 3: Use the feature importance evaluation algorithm to calculate the contribution of each data sample in the training set to the prediction result, and obtain the global importance score of each data sample; Step 4: Based on the global importance score, select an important feature subset using a feature selection method; Step 5: Train an agent model with inherent interpretability using the aforementioned important feature subset, and perform performance verification; Step 6: Obtain the sample to be predicted and preprocess it according to the method in Step 1 to obtain the preprocessed sample to be predicted. Step 7: The preprocessed samples to be predicted are input in parallel into the original prediction model and the surrogate model; the original prediction model outputs the first prediction result, and the surrogate model outputs the second prediction result and interpretable data; Step 8: Output the final predicted value based on the prediction results of the original prediction model, and at the same time output the interpretable data generated by the proxy model.

2. The power fault prediction method based on training and inference of a big data model with interpretable important features as described in claim 1, characterized in that, The preprocessing method includes: removing outlier data using the IQR method, filling missing values ​​with the mean or mode according to feature type, encoding unordered / ordered categorical features with one-hot / label encoding respectively, and normalizing continuous features with Z-Score or Min-Max.

3. The power fault prediction method based on training and inference of a big data model with interpretable important features as described in claim 1, characterized in that, The candidate prediction models include XGBoost, LSTM, Transformer, CNN, and Logistic models. The performance evaluation method for selecting the best model is as follows: for classification tasks, precision, recall, and F1 score are used; for regression tasks, root mean square error, mean absolute error, and coefficient of determination are used.

4. The power fault prediction method based on training and inference of a big data model with interpretable important features as described in claim 1, characterized in that, The global importance score uses the SHAP algorithm, and the expression is as follows: Where f is the target feature, n is the total number of training samples, and |SHAP i,f | represents the SHAP value of feature f in the i-th sample, and MA-SHAP(f) is the mean absolute error of the SHAP value; The global importance score is obtained by calculating the mean absolute error of the SHAP value of each feature across all samples.

5. The power fault prediction method based on training and inference of a big data model with interpretable important features as described in claim 1, characterized in that, The feature selection method includes any one of Top-K selection, threshold selection, and elbow rule.

6. The power fault prediction method based on training and inference of a big data model with interpretable important features as described in claim 5, characterized in that, The elbow rule selects the smallest number of features with a cumulative contribution of ≥90% as an important feature subset by plotting a line graph of the number of features versus the cumulative contribution.

7. The power fault prediction method based on training and inference of a big data model with interpretable important features as described in claim 1, characterized in that, The proxy models include logistic regression, linear regression, shallow decision tree, and generalized additive models.

8. The power fault prediction method based on training and inference of a big data model with interpretable important features as described in claim 7, characterized in that, When selecting the optimal surrogate model, for classification models, the difference in accuracy between the surrogate model and the original prediction model should be ≤3%; for regression models, the difference in mean absolute error between the surrogate model and the original prediction model should be ≤5%.

9. The power fault prediction method based on training and inference of a big data model with interpretable important features as described in claim 1, characterized in that, The interpretable data includes: the specific contribution value, contribution direction, and visualization charts for each important feature.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 9.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 9.

Images