A transformer area total list fault identification method and system based on spatial correlation and improved stacking strategy, an electronic device, and a computer readable storage medium

Abstract

This invention relates to the field of power system fault diagnosis technology, specifically disclosing a method, system, electronic device, and computer-readable storage medium for identifying transformer substation faults based on spatial correlation and an improved Stacking strategy. The method includes: collecting electrical measurement data of the target transformer substation and multivariate data of adjacent substations; using principal component analysis to reduce the dimensionality of adjacent substation data and aggregate it into one-dimensional exogenous features, which are then concatenated with the target features to form a joint input dataset; constructing an improved Stacking model containing a base learner layer, an attention weighting module, and a meta-learner layer; learning adaptive weights for the samples output by each base learner on the validation set through the attention weighting module, and performing weighted fusion of the preliminary prediction results; inputting the fused representation into the LightGBM model of the meta-learner layer, and outputting the fault classification result through nonlinear combination. This method is applicable to real-time monitoring of multiple types of faults in transformer substation meters.

Description

Technical Field

[0001] This invention relates to the field of power system automation and fault diagnosis technology, and in particular to a method, system, electronic device and computer-readable storage medium for identifying faults in transformer substations based on spatial correlation and an improved stacking strategy. Background Technology

[0002] As a key device in low-voltage distribution networks used for electricity metering and consumption recording, the main distribution meter directly affects the power supply company's electricity management, line loss analysis, and anomaly monitoring. With the increasingly complex operating environment of low-voltage distribution networks, the electricity load in distribution areas exhibits stronger fluctuations and nonlinear characteristics, posing challenges to traditional fault identification methods in terms of accuracy and real-time performance.

[0003] Existing methods for diagnosing transformer substation faults mainly fall into two categories: those based on physical models and those based on data-driven models. Physical model methods rely on precise equipment parameters and grid topology, and are poorly adaptable to dynamically changing load scenarios. While data-driven models can uncover nonlinear relationships in massive amounts of data, they are prone to misjudgment or overfitting when fault samples are scarce or when noise interference is present.

[0004] Furthermore, existing methods typically analyze only data from a single transformer area, neglecting the spatial correlation between adjacent transformer areas, or they rely solely on simple cascading or weighted strategies to fuse information, making it difficult to balance the accuracy and robustness of fault identification. Therefore, there is an urgent need for an automatic fault identification method for transformer area master tables that can fully utilize the spatial correlation between transformer areas, combine multi-model integration strategies, and consider both periodic and transient characteristics, in order to improve the identification accuracy and stability under complex load environments. Summary of the Invention

[0005] This invention provides a method, system, electronic device, and computer-readable storage medium for fault identification of transformer area master table based on spatial correlation and improved stacking strategy. It aims to solve the problems of existing technologies that rely only on single transformer area data, ignore spatial correlation, and make it difficult for a single model to take into account both periodic and transient characteristics.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for fault identification of transformer area master tables based on spatial correlation and improved stacking strategy includes the following steps: Step 1: Data Acquisition and Feature Extraction Obtain electrical measurement data from the target transformer area's master meter and multivariate measurement data from adjacent transformer areas; extract features from the electrical measurement data of the target transformer area's master meter to obtain target features; perform dimensionality reduction processing on the multivariate measurement data of adjacent transformer areas to obtain exogenous features; concatenate the exogenous features with the target features to construct a joint input dataset.

[0007] In one implementation, the adjacent transformer substations are determined based on their spatial adjacency with the target substation, power supply topology, and / or correlation with historical measurement data; dimensionality reduction of the multivariate measurement data of adjacent transformer substations includes extracting principal component scores as exogenous features using principal component analysis.

[0008] Step 2: Construct a base learner layer with complementary properties The joint input data is fed into the base learner layer, and multiple base learners are used to predict the fault category of the joint input data to obtain multiple preliminary prediction results. In one embodiment, the base learner layer of the Stacking strategy selects three deep learning models that have significant complementarity in characterizing temporal features: TimesNet, TimeMixer, and InceptionTime.

[0009] The TimesNet model utilizes its multi-period basis expansion mechanism to focus on capturing long-term periodic dependencies and phase drift characteristics in electrical data. The TimeMixer model utilizes its time-channel hybrid structure to focus on learning the nonlinear coupling relationships between multiple variables and the global dependency features across variables. The InceptionTime model utilizes its multi-branch convolutional structure to focus on identifying short-term transient abrupt changes such as spikes and drops in electrical data.

[0010] By working together, the three models described above can comprehensively extract fault features from three dimensions: periodicity, globality, and locality, and output preliminary predicted probabilities for each category.

[0011] Step 3: Adaptive Weighted Fusion of Samples Based on Attention Mechanism An attention weighting module is introduced between the base learner layer and the meta learner layer. This module calculates the contribution of each base learner's output based on the validation set, dynamically learns adaptive weights at the sample level, and performs weighted fusion on the preliminary prediction results of the base learners to generate a fused representation, thereby enhancing the model's adaptability to different sample features.

[0012] Step 4: Nonlinear correction of the meta-learner based on LightGBM The second layer of the Stacking ensemble model is constructed using LightGBM as the meta-learner. The fused representation obtained in step three is used as input, and LightGBM's nonlinear fitting capability is utilized to perform secondary combination and error correction on the weighted probability distribution. LightGBM learns the residual relationship between the prediction results of the base learner and the true labels, and outputs the final fault classification and identification results.

[0013] Preferably, the fault classification and identification results include at least the following four categories of labels: normal state, B-phase short circuit fault, C-phase short circuit fault, and three-phase short circuit fault.

[0014] The present invention also provides a corresponding fault identification system for transformer area master table, including a data preprocessing module, a base learner module, an attention weighting module, a meta-learner module and an output module. Each module respectively realizes the functions of data acquisition and feature construction, base model training, sample adaptive weighting, nonlinear combination and error correction and fault diagnosis result output.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the above-described method.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0017] Compared with the prior art, the present invention has the following advantages: 1. Introducing spatial correlation features: By using the principal component features of adjacent transformer areas as exogenous variables, the spatial correlation between transformer areas is effectively utilized, significantly improving the accuracy of fault identification.

[0018] 2. Improved Stacking Strategy for Enhanced Robustness: The attention weighting module dynamically adjusts the output weights of the base learners and combines them with the LightGBM meta-learner for nonlinear combination and error correction. Compared with a single model, it exhibits stronger anti-interference ability and generalization ability in multi-level noise environments.

[0019] 3. Complementary advantages of multiple models: TimesNet is used for periodic feature recognition, TimeMixer is used for global dependency modeling, and InceptionTime is used for transient feature capture. The combination of the three enables comprehensive recognition of fault features in the transformer substation.

[0020] Through the above technical solution, the present invention can achieve high-precision and robust fault diagnosis of transformer area meters under complex load environments, while taking into account the identification of periodic and transient characteristics, thus improving the applicability, robustness and scalability of the fault identification method. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention; Detailed Implementation To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0022] A method for fault identification of transformer area master tables based on spatial correlation and improved stacking strategy, the method comprising the following steps: S1: Obtain the electrical measurement data of the target transformer area's master meter, and obtain the multivariate measurement data of the adjacent transformer areas corresponding to the target transformer area; extract features from the electrical measurement data of the target transformer area to obtain target transformer area features; perform dimensionality reduction processing on the multivariate measurement data of the adjacent transformer areas to obtain exogenous features; concatenate the target transformer area features with the exogenous features to construct joint input data; S2: Input the joint input data into the base learner layer, and use multiple base learners to predict the fault category of the joint input data to obtain multiple preliminary prediction output results; S3: An attention weighting module is introduced between the base learner layer and the meta learner layer. The attention weighting module is used to learn the sample adaptive weights of the output of each base learner on the validation set. The preliminary prediction output results are weighted and fused to obtain the fused representation. S4: Input the fused representation into the meta-learner, use the meta-learner to perform combined correction on the weighted fused result, and output the fault identification result of the target station area summary table.

[0023] The electrical measurement data of the target transformer area collected in step S1 includes at least one of the following: three-phase current, three-phase voltage, active power, reactive power, power factor, voltage imbalance, current imbalance, and load rate.

[0024] In step S1, the dimensionality reduction processing of the multivariate measurement data of adjacent transformer substations includes: using principal component analysis to perform dimensionality reduction and aggregation on the multivariate measurement data of adjacent transformer substations, and extracting the principal component scores as the exogenous features.

[0025] The adjacent transformer areas are determined based on their spatial adjacency with the target transformer area, power supply topology, and correlation with historical measurement data.

[0026] In step S2, the multiple base learners include the TimesNet model, the TimeMixer model, and the InceptionTime model, which are complementary in feature characterization.

[0027] In step S2, each base learner outputs the prediction result of the target station area belonging to each fault category; In step S3, the attention weighting module calculates the corresponding weights based on the prediction results of each base learner, and uses the weights to perform a weighted summation of the prediction results of each base learner to generate the fusion representation.

[0028] The meta-learner in step S4 is the LightGBM model; The LightGBM model is used to output the final identification result of the fault category based on the fused representation.

[0029] As a preferred embodiment of the present invention, the present invention also provides a fault identification system for transformer area master tables based on spatial correlation and improved stacking strategy, the system comprising: The data processing module is used to acquire electrical measurement data of the target transformer area's master meter and multivariate measurement data of adjacent transformer areas, extract features from the electrical measurement data of the target transformer area, perform dimensionality reduction processing on the multivariate measurement data of adjacent transformer areas, and construct joint input data. The base learner module is used to input the joint input data into multiple base learners respectively to obtain multiple preliminary prediction output results; The attention weighting module is used to calculate the sample adaptive weights corresponding to each base learner based on the multiple preliminary prediction outputs, and to perform weighted fusion on the multiple preliminary prediction outputs to obtain a fused representation; The meta-learner module is used to input the fused representation into the meta-learner and output the fault identification results of the target station area summary table.

[0030] As a preferred embodiment of the present invention, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the steps of the method described above.

[0031] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, performs the steps of the method described above.

[0032] The following embodiment provides a method for identifying transformer substation master meter faults based on spatial correlation and an improved Stacking strategy. This method incorporates spatial features of adjacent transformer substations and combines them with an improved Stacking ensemble learning strategy to achieve high-precision automatic diagnosis of transformer substation master meter faults. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a schematic diagram illustrating the overall process and model architecture of the fault identification method described in this embodiment. As shown in the figure, the method mainly includes the following steps: Step S1: Data Acquisition and Target Feature Extraction First, historical measurement data of the target distribution transformer area's main meter is obtained from the electricity consumption information collection system. In this embodiment, the main meter data of a 10kV distribution transformer area in a certain city is selected, with a time granularity of 15 minutes and a time span from July 2019 to August 2020.

[0033] Extract multidimensional electrical measurement features of the target transformer area from the raw data and use them as model input: 1. Basic Measurement Values: Three-phase current ( ), three-phase voltage ( ), three-phase active power ( ), three-phase reactive power ( The total power, power factor, voltage imbalance, current imbalance, and load factor are used as input characteristics.

[0034] 2. Comprehensive calculation indicators: Total active power Total reactive power ; Power factor The calculation method is as follows: Voltage imbalance The calculation method is as follows: in, This is the average voltage of the three phases; Current imbalance The calculation method is as follows: in, This is the average current of the three phases; load rate The calculation method is as follows: in This refers to the rated power of the transformer.

[0035] To eliminate the impact of differences in feature dimensions on model training, the Min-Max normalization method is used to map all features to... Interval. The normalization formula is: in, The original value, The normalized value. and These are the minimum and maximum values ​​of the feature in the sample set, respectively.

[0036] Step S2: Introduction of adjacent station features based on principal component analysis Considering the similarities between adjacent transformer substations in terms of load characteristics and climate environment, this embodiment introduces data from adjacent transformer substations to enhance the robustness of the features.

[0037] 1. Relevance screening: Based on the spatial location, power supply relationship and historical operation data of the target transformer area, at least one adjacent transformer area is selected as a reference from the candidate transformer areas.

[0038] In this embodiment, correlation analysis can be further performed on the three-phase voltage and current measurement data of the target transformer area and the candidate transformer area to help determine the degree of correlation between them. The correlation analysis can use one or more of the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient, and calculate the significance probability (p-value). In this embodiment, the significance level is set to 0.05, and only transformers with a comprehensive correlation coefficient greater than 0.6 are retained. The data of adjacent transformer substations are used as a valid reference.

[0039] 2. Feature Dimensionality Reduction: For the selected multidimensional measurement data (such as voltage, current, power, etc.) of adjacent transformer substations, Principal Component Analysis (PCA) algorithm is used for dimensionality reduction. By calculating the eigenvalues ​​and eigenvectors of the covariance matrix, the first principal component is extracted as a one-dimensional exogenous feature representing the overall operating status of adjacent transformer substations.

[0040] 3. Feature Fusion: The obtained one-dimensional exogenous features are concatenated and aligned with the target area features extracted in step S1 to construct a 16-dimensional joint input feature matrix containing spatiotemporal correlation information, which is used as the input of the subsequent model.

[0041] Step S3: Build an improved Stacking integration model This embodiment constructs an improved Stacking fault identification model that includes an attention weighting mechanism. The model consists of a base learner layer, an attention weighting module, and a meta-learner layer.

[0042] 1. Construction of the base learner layer In the first layer of Stacking, three deep learning models that are complementary in temporal feature extraction—TimesNet, TimeMixer, and InceptionTime—are selected as base learners: The TimesNet model utilizes its multi-period basis expansion mechanism to capture the long-term daily / weekly periodic dependence and phase drift in electrical data, and has strong discriminative power for fault samples accompanied by periodic disturbances.

[0043] The TimeMixer model utilizes its inherent gated time-channel hybrid structure to focus on learning the nonlinear coupling relationships between multiple variables, making it suitable for cross-measurement correlation analysis with complex feature dimensions.

[0044] The InceptionTime model utilizes its multi-branch, one-dimensional convolutional structure with different receptive fields to extract short-term transient shape features such as spikes and sudden drops, while maintaining robustness to noise and time shift.

[0045] In the training configuration of this embodiment, the sequence length of each model is set to 15, the model dimension is set to 32, and the Dropout ratio is set to 0.1. Each base learner is trained on the input data and outputs the preliminary predicted probability of belonging to each category (normal, B-phase short circuit, C-phase short circuit, three-phase short circuit).

[0046] 2. Attention-based adaptive weighting of samples To address the issue of fixed weights in traditional Stacking, an Attention weighting module is introduced at the output of the base learners: the output of each base learner on the validation set is used as input.

[0047] The weight coefficients of each model's output are calculated using a fully connected layer. This weight reflects the contribution of different models to different samples.

[0048] The outputs of the base learners are weighted and summed to generate a fusion representation. : in, For the first The output of each base learner. This step enables dynamic adjustment of model weights based on sample characteristics, enhancing the flexibility of the ensemble model.

[0049] 3. Nonlinear correction of the meta-learner layer In the second layer of the stacking algorithm, LightGBM (Light Gradient Boosting Machine) is selected as the meta-learner. The fused representation output by the attention-weighted module is used as the input feature of LightGBM. LightGBM constructs multiple decision trees and performs nonlinear combination and residual learning on the prediction results of the base learners, thereby correcting the possible biases of a single model and outputting the final fault classification result.

[0050] Step S4: Model Training and Fault Identification 1. Sample Construction: Extract 500 independent samples from each of the four types of faults (normal, B-phase short circuit, C-phase short circuit, and abnormal drop in three-phase power) to construct a dataset containing 2000 samples, and divide it into training set and test set in a 6:4 ratio.

[0051] 2. Model Training: First, TimeNet, TimeMixer, and InceptionTime models are trained using the training set. Second, the Attention weighting module is trained using the output of the base learner on the validation set to obtain the adaptive weights of each model. Finally, the weighted output is used as a new feature to train the LightGBM meta-learner.

[0052] 3. Fault Identification: After the real-time measurement data of the test area is processed and feature constructed in the same way, it is input into the trained model. LightGBM outputs the fault category label (such as "B phase short circuit") and confidence level of the sample.

[0053] Experimental verification: To verify the effectiveness of the method in this embodiment, a comparative experiment was conducted.

[0054] Table 1. Comparison of classification accuracy of different models (without spatial features) Table 2. Accuracy Comparison After Introducing Adjacent Transit Area Features Table 3. Model robustness test under 15% noise interference As shown in Tables 1 and 2, compared to the single model, the improved Stacking strategy proposed in this embodiment significantly improves all indicators after introducing spatial correlation features, with an F1 score of 98.79%. As shown in Table 3, even with 15% Gaussian noise, the method in this embodiment still maintains a high accuracy, demonstrating its excellent noise resistance and generalization ability.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A transformer area total list fault identification method based on spatial correlation and improved stacking strategy, characterized in that, The method includes the following steps: S1: Obtain the electrical measurement data of the target transformer area's master meter, and obtain the multivariate measurement data of the adjacent transformer areas corresponding to the target transformer area; extract features from the electrical measurement data of the target transformer area to obtain target transformer area features; perform dimensionality reduction processing on the multivariate measurement data of the adjacent transformer areas to obtain exogenous features; concatenate the target transformer area features with the exogenous features to construct joint input data; S2: Input the joint input data into the base learner layer, and use multiple base learners to predict the fault category of the joint input data to obtain multiple preliminary prediction output results; S3: An attention weighting module is introduced between the base learner layer and the meta learner layer. The attention weighting module is used to learn the sample adaptive weights of the output of each base learner on the validation set. The preliminary prediction output results are weighted and fused to obtain the fused representation. S4: Input the fused representation into the meta-learner, use the meta-learner to perform combined correction on the weighted fused result, and output the fault identification result of the target station area summary table.

2. The method according to claim 1, characterized in that, The electrical measurement data of the target transformer area collected in step S1 includes at least one of the following: three-phase current, three-phase voltage, active power, reactive power, power factor, voltage imbalance, current imbalance, and load rate.

3. The method according to claim 1, characterized in that, In step S1, the dimensionality reduction processing of the multivariate measurement data of adjacent transformer substations includes: using principal component analysis to perform dimensionality reduction and aggregation on the multivariate measurement data of adjacent transformer substations, and extracting the principal component scores as the exogenous features.

4. The method according to claim 1, characterized in that, The adjacent transformer areas are determined based on their spatial adjacency with the target transformer area, power supply topology, and correlation with historical measurement data.

5. The method according to claim 1, characterized in that, In step S2, the multiple base learners include the TimesNet model, the TimeMixer model, and the InceptionTime model, which are complementary in feature characterization.

6. The method according to claim 1, characterized in that, In step S2, each base learner outputs the prediction result of the target station area belonging to each fault category; In step S3, the attention weighting module calculates the corresponding weights based on the prediction results of each base learner, and uses the weights to perform a weighted summation of the prediction results of each base learner to generate the fusion representation.

7. The method according to claim 1, characterized in that, The meta-learner in step S4 is the LightGBM model; The LightGBM model is used to output the final identification result of the fault category based on the fused representation.

8. A system employing the fault identification method for transformer area master tables based on spatial correlation and improved stacking strategy as described in any one of claims 1-7, characterized in that, The system includes: The data processing module is used to acquire electrical measurement data of the target transformer area's master meter and multivariate measurement data of adjacent transformer areas, extract features from the electrical measurement data of the target transformer area, perform dimensionality reduction processing on the multivariate measurement data of adjacent transformer areas, and construct joint input data. The base learner module is used to input the joint input data into multiple base learners respectively to obtain multiple preliminary prediction output results; The attention weighting module is used to calculate the sample adaptive weights corresponding to each base learner based on the multiple preliminary prediction outputs, and to perform weighted fusion on the multiple preliminary prediction outputs to obtain a fused representation; The meta-learner module is used to input the fused representation into the meta-learner and output the fault identification results of the target station area summary table.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 7.

10. 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 steps of the method as described in any one of claims 1 to 7.

Images