Transformer oil dissolved gas prediction method, device, equipment and storage medium
By constructing a heterogeneous network and using a causal importance index to screen key features, and combining this with dual-channel co-evolutionary optimization, the problem of insufficient characterization of multi-feature interactions in the prediction of dissolved gases in transformer oil is solved, improving prediction accuracy and model interpretability, and supporting transformer fault diagnosis and decision analysis.
Patent Information
- Application Number
- CN202611314378.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to characterize the complex high-order interactions between multiple features in predicting dissolved gases in transformer oil, leading to the omission of key combined features. Furthermore, the model's prediction accuracy and interpretability are insufficient, and the lack of a collaborative iterative mechanism between feature selection and model parameter optimization makes it difficult to obtain the globally optimal configuration.
Heterogeneous networks are constructed for feature interaction modeling. Key features are selected through causal importance index and co-evolution optimization is performed in a dual-channel co-evolutionary architecture. Combined with non-dominated ranking and comprehensive evaluation algorithms, a target trade-off model is obtained for gas prediction.
It improves the accuracy and robustness of dissolved gas prediction in transformer oil, enhances the interpretability of the model, provides clear feature contribution quantification, and supports transformer fault diagnosis and decision analysis.
Smart Images

Figure CN123120116A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of gas prediction technology in transformer oil, and in particular to a method, apparatus, equipment and storage medium for predicting dissolved gases in transformer oil. Background Technology
[0002] Power transformers are core equipment in power systems, and their operational stability directly determines the safe and reliable operation of the power grid. Dissolved gas analysis in oil is an important technical means to determine the degree of insulation aging in transformers and to identify early latent faults.
[0003] Current mainstream methods for predicting dissolved gases in transformer oil mostly rely on pairwise correlation analysis between features and target variables. This approach can only measure the linear or nonlinear correlation between a single feature and gas concentration, failing to characterize complex high-order interactions between multiple features. It not only easily overlooks key combined features with joint effects but also introduces numerous spurious correlation features, leading to decreased prediction model accuracy and a lack of rational decision-making basis. Furthermore, existing technologies generally separate feature selection and model parameter optimization into independent processes, lacking bidirectional feedback and collaborative iteration mechanisms. This results in a lack of focus in the optimization process, low overall optimization efficiency, and difficulty in obtaining globally optimal model configurations. In addition, most prediction models prioritize fitting accuracy while neglecting interpretability, failing to quantify the actual contribution of each feature to the prediction results, and thus struggling to support fault diagnosis and decision analysis in engineering fields. To address these technical shortcomings, there is an urgent need to construct a prediction method that integrates feature interaction modeling, causal feature selection, and collaborative optimization to improve the accuracy, robustness, and practical value of dissolved gas prediction in transformer oil. Summary of the Invention
[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, apparatus, equipment and storage medium for predicting dissolved gases in transformer oil, which aims to achieve efficient feature screening and collaborative optimization, and improve the accuracy of gas prediction and model interpretability.
[0005] The first aspect of this invention provides a method for predicting dissolved gases in transformer oil, comprising: acquiring an original feature set; constructing a heterogeneous network based on the original feature set; performing feature interaction modeling on the heterogeneous network to obtain target node embeddings; calculating the causal importance index of all features in the original feature set based on the target node embeddings, and selecting multiple key features from the original feature set based on the causal importance indexes; inputting each key feature into a pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization to obtain multi-objective optimization results; performing non-dominated sorting on the multi-objective optimization results to obtain a frontier non-dominated solution set, and selecting a target trade-off model from the frontier non-dominated solution set; and inputting each key feature into the target trade-off model for gas prediction to obtain gas concentration data.
[0006] Optionally, in a first implementation of the first aspect of the present invention, obtaining the original feature set includes: obtaining a historical dataset; performing missing value interpolation on the historical dataset using the K-nearest neighbor algorithm to obtain an interpolated dataset; performing outlier correction on the interpolated dataset using the 3σ criterion algorithm to obtain a corrected dataset; performing normalization on the corrected dataset using a normalization algorithm to obtain a normalized dataset; and performing feature extraction on the normalized dataset using a feature vectorization algorithm to obtain the original feature set.
[0007] Optionally, in a second implementation of the first aspect of the present invention, the step of constructing a heterogeneous network based on the original feature set includes: acquiring multiple gas nodes and using each feature in the original feature set as a feature node; constructing a first feature edge set based on the gas nodes and the feature nodes, and constructing a second feature edge set based on the feature nodes; calculating the edge weights of the first feature edge set using the Pearson correlation coefficient algorithm to obtain a first weight set; calculating the edge weights of the second feature edge set using the mutual information algorithm to obtain a second weight set; and constructing a weight matrix of the heterogeneous network based on the first weight set and the second weight set to complete the construction of the heterogeneous network.
[0008] Optionally, in a third implementation of the first aspect of the present invention, the step of performing feature interaction modeling processing on the heterogeneous network to obtain the target node embedding includes: obtaining a preset projection matrix; performing feature projection processing on each node in the heterogeneous network using the projection matrix to obtain an initial node embedding; calculating an attention score on the initial node embedding using an attention single-layer network to obtain an original attention score; normalizing the original attention score using a softmax function to obtain a standardized attention coefficient; performing weighted aggregation processing on the neighbor node embeddings of each node in the heterogeneous network based on the standardized attention coefficient to obtain an aggregated node embedding; and iteratively updating the aggregated node embedding using a multi-layer graph attention network to obtain the target node embedding.
[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the step of calculating the causal importance index of all features in the original feature set based on the target node embedding, and selecting multiple key features from the original feature set based on the causal importance index, includes: inputting the target node embedding and the original feature set into a pre-trained gradient importance calculation model to calculate gradient importance, obtaining multiple gradient importance values; performing counterfactual intervention processing on all features in the original feature set using a counterfactual reasoning strategy to obtain a counterfactual importance value corresponding to each gradient importance value; using a preset balance coefficient to weight and fuse each gradient importance value and the corresponding counterfactual importance value to obtain multiple causal importance indices; and sorting the multiple causal importance indices, selecting a preset number of features that rank highest as the key features.
[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the dual-channel co-evolutionary architecture includes a population initialization module, a dual-channel parallel optimization module, and a co-evaluation iteration module, wherein the population initialization module, the dual-channel parallel optimization module, and the co-evaluation iteration module are connected sequentially; the step of inputting each key feature into the pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization processing to obtain a multi-objective optimization result includes: performing double-layer encoding processing on each key feature based on the population initialization module to obtain an initial optimized population; performing parallel optimization processing on the initial optimized population based on the dual-channel parallel optimization module to obtain an optimized feature parameter combination; and performing iterative update processing on the optimized feature parameter combination based on the co-evaluation iteration module to obtain the multi-objective optimization result.
[0011] Optionally, in a sixth implementation of the first aspect of the present invention, the step of performing non-dominated ranking processing on the multi-objective optimization results to obtain a frontier non-dominated solution set, and selecting an objective trade-off model from the frontier non-dominated solution set, includes: obtaining a preset non-dominated ranking rule and all optimized individuals in the multi-objective optimization results; using the non-dominated ranking rule to perform dominance relationship judgment and screening processing on all optimized individuals to obtain the frontier non-dominated solution set; using a comprehensive evaluation algorithm to calculate the comprehensive performance score of all optimized individuals in the frontier non-dominated solution set, and selecting the model corresponding to the optimized individual with the highest comprehensive performance score as the objective trade-off model.
[0012] A second aspect of the present invention provides a device for predicting dissolved gases in transformer oil, comprising: a heterogeneous network construction module for acquiring an original feature set and constructing a heterogeneous network based on the original feature set; a feature interaction modeling module for performing feature interaction modeling processing on the heterogeneous network to obtain target node embeddings; a key feature acquisition module for calculating the causal importance index of all features in the original feature set based on the target node embeddings, and selecting multiple key features from the original feature set based on the causal importance indexes; an optimization result acquisition module for inputting each of the key features into a pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization processing to obtain multi-objective optimization results; a target model acquisition module for performing non-dominated sorting processing on the multi-objective optimization results to obtain a frontier non-dominated solution set, and selecting a target trade-off model from the frontier non-dominated solution set; and a gas prediction module for inputting each of the key features into the target trade-off model for gas prediction to obtain gas concentration data.
[0013] A third aspect of the present invention provides a device for predicting dissolved gases in transformer oil, the device comprising: a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the device for predicting dissolved gases in transformer oil to perform the steps of the method for predicting dissolved gases in transformer oil as described above.
[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the method for predicting dissolved gases in transformer oil as described in any of the preceding claims.
[0015] In the technical solution of this invention, an original feature set is first obtained, a heterogeneous network is constructed based on the original feature set, and feature interaction modeling is performed on the heterogeneous network to obtain the target node embedding. Based on the target node embedding, the causal importance index of all features in the original feature set is calculated, and multiple key features are selected from the original feature set based on the causal importance index. Each key feature is input into a pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization to obtain multi-objective optimization results. Then, the multi-objective optimization results are subjected to non-dominated ranking processing to obtain a frontier non-dominated solution set, and a target trade-off model is selected from the frontier non-dominated solution set. Finally, each key feature is input into the target trade-off model for gas prediction to obtain gas concentration data. The aim is to achieve efficient feature selection and co-optimization, and improve the accuracy of gas prediction and the interpretability of the model. Attached Figure Description
[0016] Figure 1 The flowchart of the method for predicting dissolved gases in transformer oil provided in the embodiments of the present invention is shown below. Figure 2 This is a schematic diagram of the structure of the dissolved gas prediction device in transformer oil provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a dissolved gas prediction device in transformer oil provided in an embodiment of the present invention. Detailed Implementation
[0017] This invention provides a method, apparatus, device, and storage medium for predicting dissolved gases in transformer oil. In this invention, the terms "first," "second," "third," "fourth," etc. (if applicable) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.
[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the method for predicting dissolved gases in transformer oil according to the present invention includes: 101. Obtain the original feature set and construct a heterogeneous network based on the original feature set; In this embodiment, the process of obtaining the original feature set first involves acquiring a historical dataset. The K-nearest neighbor algorithm is then used to interpolate missing values in the historical dataset, resulting in an interpolated dataset. The mean of the neighborhood feature values from the K-nearest neighbor algorithm is used to fill in the missing data, ensuring data integrity. Next, the 3σ criterion algorithm is used to correct outliers in the interpolated dataset, resulting in a corrected dataset. Thresholds for the feature mean and standard deviation are used to correct and remove outliers, improving data accuracy. Then, a normalization algorithm is used to normalize the corrected dataset, resulting in a normalized dataset. This eliminates dimensional differences between different features, achieving data scale uniformity and ensuring the stability of subsequent calculations. Finally, a feature vectorization algorithm is used to extract features from the normalized dataset, resulting in the original feature set. The original feature set is a standardized feature set composed of various effective features such as historical data on dissolved gases in transformer oil and operating status parameters, providing high-quality basic data support for subsequent network construction.
[0019] In this embodiment, after obtaining the original feature set, a heterogeneous network is constructed based on the original feature set. First, multiple gas nodes are obtained, and each feature in the original feature set is used as a feature node to complete the node system construction of the heterogeneous network. Then, a first feature edge set is constructed based on the gas nodes, and a second feature edge set is constructed based on the feature nodes to complete the edge structure construction of the heterogeneous network. Next, the edge weights of the first feature edge set are calculated using the Pearson correlation coefficient algorithm to obtain the first weight set. The linear correlation between features and gases is quantified by the Pearson correlation coefficient. The edge weights of the second feature edge set are calculated using the mutual information algorithm to obtain the second weight set. The nonlinear correlation characteristics between features are characterized by the mutual information algorithm. Finally, the weight matrix of the heterogeneous network is constructed based on the first weight set and the second weight set to complete the construction of the heterogeneous network. The heterogeneous network is a heterogeneous graph structure that integrates two types of heterogeneous nodes: feature nodes and gas nodes. The edges are defined by the correlation between features and the correlation between features and gases, and the corresponding weights are defined by the edge weights. This structure can comprehensively characterize the complex correlations between features and between features and target gases.
[0020] Through multi-stage data preprocessing, data quality is improved from multiple dimensions such as completeness, accuracy, and consistency, laying a solid data foundation for subsequent modeling. At the same time, by constructing heterogeneous networks and integrating linear and nonlinear correlation measures, higher-order interaction effects between features and the correlation characteristics between features and gas are fully captured. This effectively solves the problems of insufficient feature interaction characterization and omission of key combined features caused by traditional methods that rely solely on pairwise correlation analysis. It provides accurate topological structure support for subsequent feature interaction modeling and significantly improves the accuracy, robustness, and interpretability of subsequent prediction models.
[0021] 102. Perform feature interaction modeling on the heterogeneous network to obtain the target node embedding; In this embodiment, feature interaction modeling is performed on the heterogeneous network. High-order interaction modeling is achieved using a graph attention network. First, a preset projection matrix is obtained. This matrix is then used to project the features of each node in the heterogeneous network, mapping the original features of the feature nodes and gas nodes to a unified feature space, resulting in initial node embeddings. These initial embeddings are low-dimensional feature representations of each node after projection transformation, providing a foundation for subsequent attention calculations. Next, a single-layer attention network is used to calculate attention scores on the initial node embeddings. This single-layer network quantifies the correlation between each node and its neighbors, yielding an initial attention score that reflects the strength of the correlation between each node and its neighbors. Then, a softmax function is used to normalize the initial attention scores, converting them into standardized attention coefficients ranging from 0 to 1 with a sum of 1. This standardization of weights between different nodes ensures the rationality of subsequent weighted aggregation. Based on standardized attention coefficients, the neighbor node embeddings of each node in the heterogeneous network are weighted and aggregated. Using the standardized attention coefficients as weights, the feature information of each node's neighbors is fused to obtain the aggregated node embedding. The aggregated node embedding incorporates the feature interaction information of neighbor nodes, effectively capturing the local correlation features of nodes. Finally, a multi-layer graph attention network is used to iteratively update the aggregated node embedding. Through the stacking of multi-layer graph attention networks, higher-order feature interaction information is continuously fused to obtain the target node embedding. The target node embedding is a node feature representation that, after iterative updates by the multi-layer graph attention network, incorporates higher-order interaction effects between features and correlation information between features and the gas, comprehensively characterizing the deep correlation characteristics of each node in the heterogeneous network.
[0022] By leveraging the attention mechanism of graph attention networks, the importance weights between different nodes are adaptively learned, effectively capturing complex high-order interaction effects between features. This solves the technical bottleneck of traditional methods, which can only characterize pairwise correlations and cannot explore the influence of combined features. At the same time, through multi-layer iterative updates, the interaction information of node features is continuously deepened, significantly improving the feature representation ability of node embedding. This provides a high-quality feature foundation for subsequent causal importance calculation and prediction model construction, effectively improving the prediction accuracy, robustness, and interpretability of subsequent models.
[0023] 103. Calculate the causal importance index of all features in the original feature set based on the target node embedding, and select multiple key features from the original feature set based on the causal importance index; In this embodiment, the gradient importance calculation model is pre-trained using a heterogeneous network sample dataset. The pre-training process uses a regression loss function (mean squared error, MSE) as the loss function, with the optimization objective being to minimize the error between the predicted output and the true label of the sample. The training termination condition is set as follows: the loss function converges to a preset threshold, or the number of iterations reaches a preset maximum number of iterations; training stops when either condition is met, resulting in a pre-trained gradient importance calculation model. First, the target node embedding and the original feature set are input into the pre-trained gradient importance calculation model to calculate gradient importance, obtaining multiple gradient importance values. These gradient importance values are obtained by calculating the gradient of the model output with respect to each feature input, quantifying the direct influence of features on the model output and providing a basis for feature importance evaluation. Subsequently, a counterfactual inference strategy is employed to perform counterfactual intervention on all features in the original feature set. After setting the features to counterfactual values, predictions are re-performed, and the difference in prediction output before and after the intervention is compared to obtain the counterfactual importance value corresponding to each gradient importance value. The counterfactual importance value is used to measure the true causal relationship between the feature and the prediction result, effectively distinguishing between statistical correlation and actual causality, and eliminating invalid features with only spurious correlations. Next, a pre-set balance coefficient is used to weight and fuse each gradient importance value and its corresponding counterfactual importance value, resulting in multiple causal importance indices. These causal importance indices are comprehensive quantitative indicators formed by fusing gradient importance and counterfactual importance. The balance coefficient adjusts the weight ratio of the two types of importance, taking into account both the direct contribution of features to the model output and the true causal effect of features on the prediction result. The balance coefficient is determined through cross-validation and can adapt to the feature selection needs of different application scenarios, achieving a comprehensive and accurate characterization of feature importance. Finally, multiple causal importance indices were ranked, and a predetermined number of features at the top of the ranking were selected as key features. These key features, selected after causal importance assessment, possess both high direct contribution and strong causal association, eliminating redundant and spurious features. This provides a concise and high-value feature foundation for subsequent model construction. By integrating gradient information and counterfactual reasoning to construct the causal importance index, the technical bottleneck of traditional feature selection methods that rely solely on statistical correlation and cannot distinguish between causal associations and spurious correlations was effectively solved. This avoids interference from redundant and spurious features, significantly improving the accuracy and effectiveness of feature selection. Simultaneously, it provides clear interpretability for the model, effectively improving the accuracy, robustness, and generalization ability of subsequent prediction models, reducing model complexity, and optimizing model operating efficiency.
[0024] 104. Input each of the key features into the pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization processing to obtain multi-objective optimization results; In this embodiment, the dual-channel co-evolutionary architecture is a dedicated architecture for feature selection and model parameter co-optimization built based on the optimization characteristics of key features. Its construction relies on a two-layer encoding rule, combines the optimization characteristics of genetic algorithms and particle swarm algorithms, and introduces the multi-objective optimization mechanism of non-dominated sorting genetic algorithms.
[0025] In this embodiment, each key feature is input into a pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization. First, based on the population initialization module, each key feature undergoes dual-layer encoding. The upper layer uses binary encoding to represent the feature selection state, and the lower layer uses real-number encoding to represent the model parameters. This constructs an initial optimization population containing feature selection and parameter encoding, providing basic population support for subsequent parallel optimization. Based on the dual-channel parallel optimization module, the initial optimization population undergoes parallel optimization. The feature selection channel uses a genetic algorithm to perform selection, crossover, and mutation operations, iteratively optimizing the feature subset. The parameter optimization channel uses a particle swarm optimization algorithm to iteratively optimize the model hyperparameters and network weights. The two channels run in parallel, and a bidirectional feedback mechanism enables collaborative iteration. The fitness of parameter optimization is updated with the feature subset, and the optimal parameters are used to feed back into feature selection, ultimately obtaining an optimized feature parameter combination, achieving bidirectional coupling between feature selection and parameter optimization. The collaborative evaluation iterative module iteratively updates the optimized feature parameter combination. First, a multi-objective fitness function is constructed to evaluate the performance of the optimized feature parameter combination from three dimensions: prediction accuracy, model complexity, and causal consistency. Prediction accuracy measures the accuracy of the model's predictions, model complexity characterizes the conciseness of the feature subset, and causal consistency ensures the causal validity of the selected features. Then, a non-dominated sorting genetic algorithm is used to perform non-dominated sorting and crowding calculation, selecting and retaining Pareto front individuals to complete the population update. This iterative process continues until the termination condition is met, ultimately yielding the multi-objective optimization result. The multi-objective optimization result is a Pareto optimal solution set obtained through dual-channel co-evolutionary optimization. It contains multiple sets of feature subsets and model parameter combinations that balance prediction accuracy, model complexity, and causal consistency, providing comprehensive optimization scheme support for subsequent selection of objective-weighted models. The co-evolutionary optimization process breaks through the technical bottleneck of isolated feature selection and parameter optimization in traditional methods. By achieving collaborative optimization of features and parameters through parallel iteration and bidirectional feedback, it effectively improves optimization efficiency and global optimization capability, avoiding the problem of getting trapped in local optima in traditional step-by-step optimization. At the same time, it introduces a multi-objective fitness function to balance prediction accuracy, model complexity, and causal consistency. While ensuring the model's prediction performance, it reduces model complexity and improves the model's interpretability and generalization ability. This provides efficient and reliable technical support for the optimization of the dissolved gas prediction model in transformer oil and significantly improves the engineering practicality of the prediction model.
[0026] 105. Perform non-dominated sorting on the multi-objective optimization results to obtain the frontier non-dominated solution set, and select the objective trade-off model from the frontier non-dominated solution set; In this embodiment, firstly, all optimized individuals from the preset non-dominated ranking rules and multi-objective optimization results are obtained. The multi-objective optimization results are Pareto candidate solution sets obtained after dual-channel co-evolution optimization, containing multiple sets of feature subsets and model parameter combinations that balance prediction accuracy, model complexity, and causal consistency. The non-dominated ranking rules are used to determine and filter the dominance relationships of all optimized individuals. By comparing the performance of each optimized individual pairwise in the three dimensions of prediction accuracy, model complexity, and causal consistency, optimized individuals dominated by other individuals are eliminated, and Pareto front individuals that are not dominated by any other individual are selected to form the front non-dominated solution set. The front non-dominated solution set contains multiple sets of optimal feature subsets and model parameter combinations that cannot dominate each other in the multi-objective optimization dimensions, providing comprehensive candidate schemes for the selection of objective-balancing models. A comprehensive evaluation algorithm is employed to calculate the overall performance score of all optimized individuals in the frontier non-dominated solution set. This algorithm integrates the weights of three indicators—prediction accuracy, model complexity, and causal consistency—to comprehensively quantify the performance of each optimized individual. The model corresponding to the optimized individual with the highest comprehensive performance score is then selected as the objective trade-off model. The architecture of this objective trade-off model is centered on a graph attention network, incorporating key feature inputs filtered by causal importance, and carrying the optimal hyperparameters and network weights obtained through co-evolutionary optimization. It also integrates the optimal feature subset determined by multi-objective optimization, forming an end-to-end prediction architecture that balances prediction accuracy, model complexity, and causal consistency. By effectively selecting the Pareto optimal solution set under multi-objective optimization through non-dominated ranking, the limitations of single-objective optimization are avoided. The comprehensive evaluation algorithm achieves an effective trade-off between multiple objectives, selecting the objective trade-off model with the best overall performance. This approach reduces model complexity while ensuring prediction accuracy, improving interpretability and generalization ability. It overcomes the technical bottleneck of traditional models struggling to balance accuracy and complexity, providing efficient and reliable model support for the prediction of dissolved gases in transformer oil, and significantly enhancing the engineering practicality and decision-making reference value of the prediction model.
[0027] 106. Input each of the key features into the target trade-off model to predict gas concentration data.
[0028] In this embodiment, each key feature is input into the target tradeoff model for gas prediction to obtain gas concentration data. First, the key features are standardized and adapted based on the input layer of the target tradeoff model, mapping the key features selected by causal importance to the feature space adapted to the model, ensuring the matching of input data and model structure. Then, through the graph attention network core architecture of the target tradeoff model, the high-order interaction relationship between key features is deeply modeled and feature extracted. Based on the hyperparameters and network weights obtained by co-evolution optimization, forward inference calculation is carried out to generate preliminary prediction results of dissolved gases in oil. Then, the preliminary prediction results are calibrated and corrected through the model output layer, and finally, the concentration data of various dissolved gases in oil such as hydrogen, methane, and acetylene are output, completing the entire gas prediction process. Based on the optimized architecture and parameter configuration of the target trade-off model, the model complexity is controlled while ensuring prediction accuracy. By effectively eliminating redundant information interference through the input of key features, the prediction efficiency and robustness are improved. At the same time, the model has clear interpretability, which can clearly define the contribution of each key feature to the gas concentration prediction. This provides accurate data support for transformer insulation status assessment, enables early warning of potential transformer faults, significantly improves the safety and stability of power system operation, and provides reliable technical basis for transformer operation and maintenance decisions.
[0029] In this embodiment of the invention, obtaining the original feature set includes: obtaining a historical dataset; performing missing value interpolation on the historical dataset using the K-nearest neighbor algorithm to obtain an interpolated dataset; performing outlier correction on the interpolated dataset using the 3σ criterion algorithm to obtain a corrected dataset; performing normalization on the corrected dataset using a normalization algorithm to obtain a normalized dataset; and performing feature extraction on the normalized dataset using a feature vectorization algorithm to obtain the original feature set.
[0030] In this embodiment, the historical dataset includes multiple types of basic data collected during transformer operation. Specifically, it includes historical concentration data of dissolved gases in transformer oil, covering the content values of characteristic gases such as hydrogen, methane, and acetylene. It also includes transformer operating parameter data, such as load current, voltage level, winding temperature, oil surface temperature, and tap changer position, as well as transformer equipment structural parameters and historical maintenance records. By integrating these multi-dimensional data, a comprehensive and complete historical dataset is constructed. The K-nearest neighbor algorithm is used to interpolate missing values in the historical dataset. By selecting the K nearest neighbor samples and using the mean of the nearest neighbor feature values to fill in the missing data, an interpolated dataset is obtained, effectively ensuring data integrity. The 3σ criterion algorithm is used to correct outliers in the interpolated dataset. Based on the feature mean and standard deviation, a threshold is calculated, and outliers exceeding the threshold range are corrected to the corresponding threshold boundary, eliminating interference from extreme outliers and improving data accuracy and reliability. A normalization algorithm is employed to normalize the corrected dataset. This normalization algorithm, which can be implemented using min-max normalization, maps feature values to the range of 0 to 1, eliminating dimensional and numerical scale differences between different features. This ensures the stability and convergence efficiency of subsequent model calculations, resulting in a normalized dataset. A feature vectorization algorithm is then used to extract features from the normalized dataset. This algorithm can be implemented using tensor encoding or sequence vectorization, converting the normalized structured data into a model-suitable vector form, thus completing the structured representation of the features and obtaining the original feature set. Through the collaborative efforts of missing value interpolation, outlier correction, normalization, and feature vectorization, data quality is improved in terms of completeness, accuracy, and consistency. This provides high-quality foundational data for subsequent heterogeneous network construction and feature interaction modeling, effectively avoiding the negative impact of original data quality defects on subsequent modeling accuracy. It significantly improves the predictive performance, robustness, and generalization ability of subsequent models, laying a solid data foundation for the prediction of dissolved gases in transformer oil.
[0031] In this embodiment of the invention, the construction of a heterogeneous network based on the original feature set includes: acquiring multiple gas nodes and using each feature in the original feature set as a feature node; constructing a first feature edge set based on the gas nodes and the feature nodes, and constructing a second feature edge set based on the feature nodes; calculating the edge weights of the first feature edge set using the Pearson correlation coefficient algorithm to obtain a first weight set; calculating the edge weights of the second feature edge set using the mutual information algorithm to obtain a second weight set; and constructing a weight matrix of the heterogeneous network based on the first weight set and the second weight set to complete the construction of the heterogeneous network.
[0032] In this embodiment, multiple gas nodes are acquired, and each feature in the original feature set is used as a feature node to construct a node system for a heterogeneous network. The heterogeneous network uses a heterogeneous graph. The core architecture consists of a set of nodes. The network consists of multiple feature nodes and multiple gas nodes. The feature nodes correspond to various feature dimensions in the original feature set, while the gas nodes correspond to various target gases dissolved in transformer oil. These two types of nodes together form the topological foundation of the network. A first feature edge set is constructed based on the gas nodes, representing the association edges between feature nodes and gas nodes, used to characterize the direct relationship between features and target gases. A second feature edge set is constructed based on the feature nodes, representing the association edges between different feature nodes, used to capture the intrinsic interaction relationships between features. These two types of edges together constitute the edge set of the heterogeneous network. The edge weights of the first feature edge set are calculated using the Pearson correlation coefficient algorithm, resulting in the first weight set. The corresponding calculation expression is as follows: , in, Indicates the first The feature node and the first The edge weight of the first characteristic edge between the gas nodes. Indicates the first Each feature node Indicates the first One gas node, Indicates the first The feature node and the first The covariance between the gas nodes is used to quantify the degree of linear coordinated change between the two. Indicates the first The standard deviation of each feature node Indicates the first The standard deviation of each gas node can be used to accurately quantify the linear correlation strength between the feature and the gas, assigning appropriate weights to the first feature edges. The mutual information algorithm is used to calculate the edge weights of the second feature edges, resulting in the second weight set. The corresponding calculation formula is: , in, Indicates the first The feature node and the first The edge weights of the second feature edges between the feature nodes. and They represent the first The feature node and the first The values of each feature node, Indicates the first Take each feature node value and the first Take each feature node The joint probability of the values, Indicates the first Take each feature node The marginal probability of the value Indicates the first Take each feature node The marginal probability of the value can effectively characterize the degree of nonlinear correlation between features through this formula, making up for the limitations of linear measures and providing a comprehensive weight representation for the second feature edge.
[0033] In this embodiment, a weight matrix for the heterogeneous network is constructed based on the first and second weight sets. The weights of the two types of edges are mapped to the corresponding positions of each node in the network, completing the construction of the heterogeneous network. This heterogeneous network takes feature nodes and gas nodes as its core, features the correlation between features and the correlation between features and gas as its edge structure, and uses differentiated weights as edge weights to comprehensively characterize the complex correlation relationships within features and between features and target gases. By distinguishing between the two types of correlation edges and using differentiated weight calculation methods, the linear correlation characteristics between features and gases and the nonlinear interaction effects between features are taken into account. This effectively solves the technical bottleneck that traditional single correlation measures cannot fully characterize complex correlation relationships. It can accurately capture the high-order interactions between features and the direct correlation between features and gases, providing accurate and comprehensive topological support for subsequent feature interaction modeling. At the same time, the construction of the weight matrix realizes the quantitative representation of correlation relationships, significantly improving the interpretability of the network structure. This lays a solid topological foundation for subsequent causal importance index calculation, model optimization, and gas prediction, effectively improving the accuracy, robustness, and generalization ability of subsequent prediction models, and providing reliable structural support for the accurate prediction of dissolved gases in transformer oil.
[0034] In this embodiment of the invention, the step of performing feature interaction modeling on the heterogeneous network to obtain the target node embedding includes: obtaining a preset projection matrix; using the projection matrix to perform feature projection processing on each node in the heterogeneous network to obtain an initial node embedding; using an attention single-layer network to calculate the attention score of the initial node embedding to obtain an original attention score; using a softmax function to normalize the original attention score to obtain a standardized attention coefficient; performing weighted aggregation processing on the neighbor node embeddings of each node in the heterogeneous network based on the standardized attention coefficient to obtain an aggregated node embedding; and using a multi-layer graph attention network to iteratively update the aggregated node embedding to obtain the target node embedding.
[0035] In this embodiment, the projection matrix is a pre-defined learnable weight matrix. Its dimensions are determined based on the original feature dimensions and target embedding dimensions of the heterogeneous network nodes, and it is used to map the features of heterogeneous nodes to a unified latent feature space. The projection matrix is used to perform feature projection processing on each node in the heterogeneous network. The specific calculation expression is as follows: , in, Represents a node The initial node embedding, Indicates the preset projection matrix. Represents a node The original feature vectors are used to eliminate the feature differences between heterogeneous nodes through mapping, resulting in initial node embeddings. An attention-based single-layer network is then used to calculate attention scores for the initial node embeddings. This single-layer network is a single-hidden-layer graph attention calculation unit, containing linear transformations and attention scoring structures. The specific calculation expression is as follows: , in, Represents the raw attention score. Represents the attention weight vector. Indicates the transpose symbol. Represents a node The initial node embedding, Represents a node The initial node embedding uses LeakyReLU, a first-class nonlinear activation function: when the input value is greater than 0, it directly outputs the input value; when the input value is less than or equal to 0, it outputs the input value multiplied by a pre-set small slope coefficient. This activation function preserves a small gradient for negative inputs, preventing neuron failure. This formula is used to perform preliminary quantification of the correlation strength between nodes, obtaining the original attention score. The softmax function is then used to normalize the original attention score; the specific calculation expression is as follows: , in, Represents the standardized attention coefficient. The exponential transformation representing the original attention score. Represents a node The set of neighboring nodes, Represents a node with neighboring nodes The original attention scores are normalized, and the sum of each attention coefficient is non-negative, resulting in standardized attention coefficients. When a node... The set of neighboring nodes When the set is empty, skip the softmax normalization and neighbor-weighted aggregation steps mentioned above, and directly add the nodes. The initial node embedding is used as the aggregate node embedding to avoid division by zero when the denominator is zero. The neighbor node embeddings of each node in the heterogeneous network are weighted and aggregated based on a standardized attention coefficient. The specific calculation expression is as follows: , in, Represents a node Updated aggregation node embedding, This represents the second type of nonlinear activation function: when the input value is greater than 0, it directly outputs the input value; when the input value is less than or equal to 0, it directly outputs 0. This activation function sets the negative input to zero, completing the nonlinear transformation of the feature. Represents a node With nodes Standardized attention coefficients between them This represents the preset projection matrix. Represents a node Node embedding, Represents a node The set of neighboring nodes is weighted and aggregated to fuse the feature information of the neighboring nodes, resulting in aggregated node embeddings. A multi-layer graph attention network is used to iteratively update the aggregated node embeddings. The multi-layer graph attention network is composed of multiple stacked single-layer attention networks, performing attention calculation and feature aggregation layer by layer, continuously fusing high-order neighbor interaction information, and finally obtaining the target node embedding. The target node embedding is a deep node feature representation that fuses topological associations and high-order interactions. A projection matrix is used to achieve a unified representation of heterogeneous node features. The single-layer attention network adaptively learns node association weights, and the multi-layer graph attention network is used to mine deep interaction relationships, which can accurately characterize the complex relationship between features and gas nodes, improve the expressive power and robustness of node embeddings, provide a high-quality feature foundation for subsequent causal importance index calculation, and effectively improve the model's prediction accuracy and generalization ability.
[0036] In this embodiment of the invention, the step of calculating the causal importance index of all features in the original feature set based on the target node embedding, and selecting multiple key features from the original feature set based on the causal importance index, includes: inputting the target node embedding and the original feature set into a pre-trained gradient importance calculation model to calculate gradient importance, obtaining multiple gradient importance values; performing counterfactual intervention processing on all features in the original feature set using a counterfactual reasoning strategy to obtain a counterfactual importance value corresponding to each gradient importance value; using a preset balance coefficient to weight and fuse each gradient importance value and the corresponding counterfactual importance value to obtain multiple causal importance indices; and ranking the multiple causal importance indices, selecting a preset number of features with the highest ranking as the key features.
[0037] In this embodiment, the pre-trained gradient importance calculation model is based on a graph attention network architecture. The target node is embedded in the original feature set as model input, and the attention mechanism of the graph attention network enables adaptive learning of the correlation between features. The model output is the gradient importance value corresponding to each feature. The training process uses historical dissolved gas data in transformer oil as the training set and the correlation labels between features and gas concentrations as supervision signals. The backpropagation algorithm is used to optimize the projection matrix and attention weights within the model. During training, by minimizing the error between the predicted and actual gradient importance values, the model can accurately capture the direct impact of features on the model output, thus completing the pre-training of the gradient importance calculation model.
[0038] In this embodiment, the target node embedding and the original feature set are input into the pre-trained gradient importance calculation model to calculate gradient importance, resulting in multiple gradient importance values. The specific calculation expression is as follows: , in, Indicates the first The gradient importance values of each feature This represents the model output, which is the prediction result of dissolved gas in oil, obtained by fusing the target node embedding with the original feature input. Indicates the first fusion target node after embedding The formula quantifies the direct contribution of the fused and embedded features to the prediction result by taking the absolute value of the partial derivative of the model output with respect to the feature input. The role of target node embedding has been fully integrated. and In the mapping relationship, the partial derivatives naturally reflect the feature interaction contribution brought by the embedding, providing a basis for feature importance evaluation. A counterfactual inference strategy is used to perform counterfactual intervention on all features in the original feature set. Specifically, the value of each feature is replaced with a preset counterfactual value, while preserving the structure and weights of the target node embedding. The intervened feature set is then input into the trained gradient importance calculation model to obtain the predicted output after feature intervention. The specific calculation expression is as follows: , in, Indicates the first The counterfactual importance value of each feature, This represents the model output, which is the prediction result of dissolved gas in oil, obtained by fusing the target node embedding with the original feature input. Representation of features After setting the counterfactual value, the predicted output based on the target node embedding is compared with the difference in predicted output before and after intervention to measure the true causal relationship between features and prediction results, effectively distinguishing between statistical correlation and actual causal relationship. The existence of target node embedding ensures the consistency of the model feature interaction structure before and after intervention, ensuring the effectiveness of counterfactual importance calculation. A preset balance coefficient is used to weight and fuse each gradient importance value and its corresponding counterfactual importance value to obtain multiple causal importance indices. The balance coefficient is set according to the results of cross-validation. The feature selection effect under different balance coefficients is verified, and an optimized λ value that can balance the direct contribution of features and causal effectiveness is selected. The specific calculation expression is as follows: , in, Indicates the first The causal importance index of each feature This represents the balance hyperparameter with a value between 0 and 1. The optimal value is determined by searching through the validation set, with the optimal model evaluation metric on the validation set serving as the selection criterion. Indicates the first The gradient importance values of each feature Indicates the first The counterfactual importance values of each feature are weighted and fused to achieve a comprehensive and accurate characterization of feature importance. This preserves the direct contribution of features to the model and ensures the effectiveness of the causal relationship between features and prediction results. The fusion of target node embeddings allows the causal importance index to cover both the original contribution and higher-order interaction contribution of features, thus improving the comprehensiveness of importance assessment.
[0039] In this embodiment, multiple causal importance indices are ranked, and a predetermined number of features at the top of the ranking are selected as key features. These key features are a set of core features selected after causal importance assessment that possess both high direct contribution and strong causal association, eliminating redundant and spurious features, thus providing a concise and high-value feature foundation for subsequent model construction. A pre-trained gradient importance calculation model is used to quantify the direct contribution of features embedded in the target node, relying on counterfactual reasoning to identify true causal associations. A balance coefficient is combined to achieve accurate fusion of the two types of importance, effectively solving the technical bottleneck of traditional feature selection methods that rely solely on statistical correlation and cannot distinguish between causal associations and spurious associations. Furthermore, integrating target node embedding into the end-to-end calculation allows importance assessment to cover both the original feature contribution and higher-order interaction contribution, avoiding the omission of embedding effects. This significantly improves the accuracy and effectiveness of feature selection, provides clear interpretability for the model, effectively enhances the accuracy, robustness, and generalization ability of subsequent prediction models, reduces model complexity, optimizes model operating efficiency, and lays a high-quality feature foundation for the accurate prediction of dissolved gases in transformer oil.
[0040] In this embodiment of the invention, the dual-channel co-evolutionary architecture includes a population initialization module, a dual-channel parallel optimization module, and a co-evaluation iteration module, which are sequentially connected. The step of inputting each key feature into the pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization processing to obtain a multi-objective optimization result includes: performing double-layer encoding processing on each key feature based on the population initialization module to obtain an initial optimized population; performing parallel optimization processing on the initial optimized population based on the dual-channel parallel optimization module to obtain an optimized feature parameter combination; and performing iterative update processing on the optimized feature parameter combination based on the co-evaluation iteration module to obtain the multi-objective optimization result.
[0041] In this embodiment, the dual-channel co-evolutionary architecture is a dedicated multi-objective optimization architecture for the co-optimization of key features and model parameters. It consists of a population initialization module, a dual-channel parallel optimization module, and a co-evaluation iteration module connected sequentially. The architecture is based on a two-layer encoding rule and is constructed by integrating the genetic algorithm, particle swarm optimization algorithm, and NSGA-II multi-objective optimization mechanism. During the architecture construction process, the two-layer encoding specification is first clarified. The upper layer uses binary encoding to represent the feature selection state, and the lower layer uses real number encoding to represent the model parameters. Based on this, three types of functional modules are built: the population initialization module is responsible for completing the two-layer encoding of key features and the initial population generation; the dual-channel parallel optimization module is responsible for the parallel optimization of feature subsets and model parameters; and the co-evaluation iteration module is responsible for multi-objective fitness evaluation and population iterative update. The three types of modules are connected sequentially to form a complete co-evolutionary optimization architecture, realizing parallel iteration and bidirectional coupling optimization of feature selection and parameter optimization.
[0042] In this embodiment, each key feature is input into a pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization. First, based on the population initialization module, each key feature is subjected to dual-layer encoding to obtain an initial optimized population. The dual-layer encoding consists of two layers: the upper layer is binary encoding, corresponding to the feature selection state s. i s i The value can be 0 or 1. 0 means that the corresponding feature is not selected, and 1 means that the corresponding feature is included in the feature subset. The lower layer is real number encoding, which corresponds to the model hyperparameters and network weights. Through double-layer encoding, the key features and model parameters are integrated into a unified population, completing the initialization of the initial optimization population and providing basic population support for subsequent parallel optimization.
[0043] In this embodiment, the initial optimization population is optimized in parallel based on a dual-channel parallel optimization module. The dual-channel parallel optimization module includes two parallel sub-modules: a feature selection channel and a parameter optimization channel. The feature selection channel uses a genetic algorithm to perform selection, crossover, and mutation operations to iteratively optimize the feature subset. The parameter optimization channel uses a particle swarm optimization algorithm to iteratively optimize the model hyperparameters and network weights. The two channels run in parallel and achieve collaborative iteration through a bidirectional feedback mechanism. The fitness of parameter optimization is updated with the feature subset, and the optimal parameters are used to feed back into feature selection, ultimately obtaining an optimized feature parameter combination, thus realizing the bidirectional coupling of feature selection and parameter optimization.
[0044] In this embodiment, the optimized feature parameter combination is iteratively updated based on the collaborative evaluation iteration module. First, a multi-objective fitness function is constructed to evaluate the performance of the optimized feature parameter combination from three dimensions: prediction accuracy, model complexity, and causal consistency. The prediction accuracy dimension quantifies the accuracy of the model's predictions using the root mean square error (RMSE). The calculation process involves summing the squared errors between the model's predicted and actual values across all samples, averaging the sum, and then taking the square root. A smaller final value indicates higher model prediction accuracy. The total number of samples represents the number of all oil dissolved gas monitoring samples involved in the calculation. The actual value is the actual detected concentration of dissolved gases in transformer oil, and the predicted value is the gas concentration prediction result output by the model. The model complexity dimension is calculated by statistically analyzing the selected features... The quantity dimension represents model complexity. The calculation process involves counting the number of features with a value of 1 among all key features. The smaller the final value, the lower the model complexity. The total number of key features is the total number of all features selected through causal importance screening, and the encoding value is the binary selection state of the corresponding feature. The causal consistency dimension quantifies the causal effectiveness of the selected features by calculating the average causal importance index of the selected features. The calculation process involves summing the causal importance indices of all selected features and then dividing by the total number of selected features. The larger the final value, the stronger the causal consistency of the selected features. The causal importance index is the quantified causal importance value of each key feature obtained in the previous calculation, and the encoding value is the binary selection state of the corresponding feature. The collaborative evaluation iteration module performs non-dominated sorting and crowding calculation based on the NSGA-II algorithm, conducts multi-objective evaluation on the optimized feature parameter combination, selects and retains Pareto front individuals, completes population update, and continues iterating until the termination condition is met, finally obtaining the multi-objective optimization result. The multi-objective optimization result is the Pareto optimal solution set obtained after co-evolution optimization, which includes multiple feature subsets and model parameter combinations that take into account prediction accuracy, model complexity and causal consistency, providing comprehensive optimization scheme support for the subsequent selection of objective trade-off models. By achieving a unified representation of features and parameters through dual-layer encoding, and relying on a dual-channel parallel optimization module to realize parallel iteration and bidirectional feedback of feature selection and parameter optimization, this approach breaks through the technical bottleneck of the isolation between feature selection and parameter optimization in traditional methods. It effectively improves optimization efficiency and global optimization capability, and avoids the problem of traditional step-by-step optimization easily getting trapped in local optima. At the same time, a multi-objective fitness function is introduced to balance prediction accuracy, model complexity, and causal consistency. While ensuring the model's prediction performance, it reduces model complexity and improves the model's interpretability and generalization ability. This provides efficient and reliable technical support for the optimization of the dissolved gas prediction model in transformer oil, and significantly improves the engineering practicality of the prediction model.
[0045] In this embodiment of the invention, the step of performing non-dominated ranking processing on the multi-objective optimization results to obtain a frontier non-dominated solution set, and selecting an objective trade-off model from the frontier non-dominated solution set, includes: obtaining a preset non-dominated ranking rule and all optimization individuals in the multi-objective optimization results; using the non-dominated ranking rule to perform dominance relationship judgment and screening processing on all optimization individuals to obtain the frontier non-dominated solution set; using a comprehensive evaluation algorithm to calculate the comprehensive performance score of all optimization individuals in the frontier non-dominated solution set, and selecting the model corresponding to the optimization individual with the highest comprehensive performance score as the objective trade-off model.
[0046] In this embodiment, the non-dominated ranking rule is a multi-objective individual selection criterion built upon the core logic of the NSGA-II algorithm. Its core principle is to identify the dominance relationships between optimized individuals layer by layer. First, it determines whether an individual is dominated by other individuals. Individuals not dominated are assigned to the first Pareto front. Then, these front individuals are removed from the population. The dominance relationship judgment is repeated for the remaining individuals, generating subsequent Pareto fronts in sequence. This completes the hierarchical selection of all optimized individuals, resulting in the front non-dominated solution set. The non-dominated ranking rule is used to perform dominance relationship judgment and selection on all optimized individuals. Specifically, by comparing the performance of each optimized individual in three dimensions—prediction accuracy, model complexity, and causal consistency—the superiority-inferiority relationships between individuals are clarified. Optimized individuals not dominated by any other individuals are selected to form the front non-dominated solution set. This solution set contains multiple sets of feature subsets and model parameter combinations that cannot mutually suppress each other in the multi-objective optimization dimensions, providing a high-quality candidate basis for selecting objective-weighted models.
[0047] In this embodiment, a comprehensive evaluation algorithm is used to calculate the comprehensive performance score of all optimization individuals in the frontier non-dominated solution set. The comprehensive evaluation algorithm can be implemented using a weighted summation algorithm or the TOPSIS algorithm. By integrating the weights of three types of indicators—prediction accuracy, model complexity, and causal consistency—a comprehensive quantitative score is given to each optimization individual. The indicator weights can be determined by combining the analytic hierarchy process (AHP) or the entropy weight method with expert experience. Then, the model corresponding to the optimization individual with the highest comprehensive performance score is selected as the target trade-off model. The target trade-off model is an optimization model configuration that takes into account the needs of multi-objective optimization. Its architecture integrates the optimal feature subset obtained through co-evolutionary optimization with model parameters, and has a balanced prediction accuracy, model complexity, and causal consistency. By relying on pre-defined non-dominated ranking rules, the Pareto optimal front under multi-objective optimization is accurately identified, avoiding the limitations of single-objective optimization. Through a comprehensive evaluation algorithm, the quantitative trade-offs of multiple objectives are realized, effectively solving the technical bottleneck of balancing multiple optimization objectives in traditional methods. This ensures that the selected objective trade-off model guarantees prediction accuracy while taking into account low model complexity and high causal consistency, improving the model's generalization ability and interpretability. It provides reliable model support for subsequent prediction of dissolved gas concentration in oil and accurate early warning of transformer status, significantly improving the scientificity and foresight of power system operation and maintenance decisions.
[0048] The above describes the method for predicting dissolved gases in transformer oil in embodiments of the present invention. The following describes the device for predicting dissolved gases in transformer oil in embodiments of the present invention. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the transformer oil dissolved gas prediction device of the present invention includes: Heterogeneous network construction module 201: used to obtain the original feature set and construct a heterogeneous network based on the original feature set; Feature interaction modeling module 202: used to perform feature interaction modeling processing on the heterogeneous network to obtain the target node embedding; Key feature acquisition module 203: used to calculate the causal importance index of all features in the original feature set based on the target node embedding, and to select multiple key features from the original feature set based on the causal importance index; Optimization result acquisition module 204: used to input each of the key features into a pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization processing to obtain multi-objective optimization results; Target model acquisition module 205: used to perform non-dominated sorting processing on the multi-objective optimization results to obtain the frontier non-dominated solution set, and to select the target trade-off model from the frontier non-dominated solution set; Gas prediction module 206: Used to input each of the key features into the target trade-off model to predict gas concentration data.
[0049] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.
[0050] above Figure 2 The dissolved gas prediction device in transformer oil in this embodiment of the invention is described in detail from the perspective of modular functional entities. The dissolved gas prediction device in transformer oil in this embodiment of the invention is described in detail below from the perspective of hardware processing.
[0051] Figure 3 This is a schematic diagram of a dissolved gas prediction device in transformer oil according to an embodiment of the present invention. The dissolved gas prediction device 300 can vary considerably depending on its configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the dissolved gas prediction device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the dissolved gas prediction device 300 to implement the steps of the dissolved gas prediction method in transformer oil provided in the above-described method embodiments.
[0052] The transformer oil dissolved gas prediction device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The structure of the transformer oil dissolved gas prediction device shown does not constitute a limitation on the transformer oil dissolved gas prediction device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.
[0053] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the method for predicting dissolved gases in transformer oil.
[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0055] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0056] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting dissolved gases in transformer oil, characterized in that, include: Obtain the original feature set and construct a heterogeneous network based on the original feature set; The heterogeneous network is subjected to feature interaction modeling to obtain the target node embedding; The causal importance index of all features in the original feature set is calculated based on the target node embedding, and multiple key features are selected from the original feature set based on the causal importance index. Each of the key features is input into a pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization to obtain multi-objective optimization results; The multi-objective optimization results are subjected to non-dominated sorting to obtain a frontier non-dominated solution set, and the objective trade-off model is selected from the frontier non-dominated solution set. Each of the key features is input into the target trade-off model to predict gas concentration data.
2. The method for predicting dissolved gases in transformer oil according to claim 1, characterized in that, The process of obtaining the original feature set includes: Obtain the historical dataset and use the K-nearest neighbor algorithm to interpolate the missing values in the historical dataset to obtain the interpolated dataset. The interpolation dataset is processed using the 3σ criterion algorithm to correct outliers, resulting in a corrected dataset. The corrected dataset is normalized using a normalization algorithm to obtain a normalized dataset; The normalized dataset is processed by feature vectorization algorithm to obtain the original feature set.
3. The method for predicting dissolved gases in transformer oil according to claim 1, characterized in that, The construction of a heterogeneous network based on the original feature set includes: Multiple gas nodes are obtained, and each feature in the original feature set is used as a feature node; A first feature edge set is constructed based on the gas nodes and the feature nodes, and a second feature edge set is constructed based on the feature nodes; The edge weights of the first feature edge set are calculated using the Pearson correlation coefficient algorithm to obtain the first weight set; The edge weights of the second feature edge set are calculated using the mutual information algorithm to obtain the second weight set; The weight matrix of the heterogeneous network is constructed based on the first weight set and the second weight set to complete the construction of the heterogeneous network.
4. The method for predicting dissolved gases in transformer oil according to claim 1, characterized in that, The step of performing feature interaction modeling on the heterogeneous network to obtain the target node embedding includes: Obtain a preset projection matrix, and use the projection matrix to perform feature projection processing on each node in the heterogeneous network to obtain the initial node embedding; An attention score is calculated on the initial node embedding using a single-layer attention network to obtain the original attention score. The original attention scores are normalized using the softmax function to obtain standardized attention coefficients; The neighbor node embeddings of each node in the heterogeneous network are weighted and aggregated based on the standardized attention coefficients to obtain the aggregated node embeddings. The target node embedding is obtained by iteratively updating the aggregated node embedding using a multi-layer graph attention network.
5. The method for predicting dissolved gases in transformer oil according to claim 1, characterized in that, The causal importance index of all features in the original feature set is calculated based on the target node embedding, and multiple key features are selected from the original feature set based on the causal importance index, including: The target node embedding and the original feature set are input into a pre-trained gradient importance calculation model to calculate gradient importance, resulting in multiple gradient importance values. A counterfactual reasoning strategy is used to perform counterfactual intervention processing on all features in the original feature set to obtain a counterfactual importance value corresponding to each gradient importance value; Each gradient importance value and its corresponding counterfactual importance value are weighted and fused using a preset balance coefficient to obtain multiple causal importance indices corresponding to the features; The multiple causal importance indices are sorted, and a predetermined number of features that rank highest are selected as the key features.
6. The method for predicting dissolved gases in transformer oil according to claim 1, characterized in that, The dual-channel co-evolutionary architecture includes a population initialization module, a dual-channel parallel optimization module, and a co-evaluation iteration module, which are sequentially connected. The step of inputting each key feature into the pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization processing to obtain multi-objective optimization results includes: Based on the population initialization module, each key feature is subjected to double-layer encoding to obtain an initial optimized population. The initial optimization population is subjected to parallel optimization processing based on the dual-channel parallel optimization module to obtain the optimized feature parameter combination. The optimized feature parameter combination is iteratively updated based on the collaborative evaluation iteration module to obtain the multi-objective optimization result.
7. The method for predicting dissolved gases in transformer oil according to claim 1, characterized in that, The process of performing non-dominated ranking on the multi-objective optimization results to obtain a frontier non-dominated solution set, and then selecting the objective trade-off model from the frontier non-dominated solution set, includes: Obtain the preset non-dominated ranking rule and all optimized individuals in the multi-objective optimization results; The non-dominated sorting rule is used to perform a dominance relationship judgment and screening process on all the optimized individuals to obtain the frontier non-dominated solution set; A comprehensive evaluation algorithm is used to calculate the comprehensive performance score of all optimization individuals in the frontier non-dominated solution set, and the model corresponding to the optimization individual with the highest comprehensive performance score is selected as the target trade-off model.
8. A device for predicting dissolved gases in transformer oil, characterized in that, include: Heterogeneous network construction module: used to obtain the original feature set and construct a heterogeneous network based on the original feature set; Feature interaction modeling module: used to perform feature interaction modeling processing on the heterogeneous network to obtain the target node embedding; Key feature acquisition module: used to calculate the causal importance index of all features in the original feature set based on the target node embedding, and to select multiple key features from the original feature set based on the causal importance index; The optimization result acquisition module is used to input each of the key features into a pre-constructed dual-channel co-evolutionary architecture for co-evolutionary optimization processing to obtain multi-objective optimization results. Target model acquisition module: used to perform non-dominated sorting processing on the multi-objective optimization results to obtain the frontier non-dominated solution set, and to select the target trade-off model from the frontier non-dominated solution set; Gas prediction module: used to input each of the key features into the target trade-off model to predict gas concentration data.
9. A device for predicting dissolved gases in transformer oil, characterized in that, The dissolved gas prediction device in transformer oil includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the transformer oil dissolved gas prediction device to perform the various steps of the transformer oil dissolved gas prediction method as claimed in any one of claims 1-7.
10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the method for predicting dissolved gases in transformer oil as described in any one of claims 1-7.
