Transformer fault evolution prediction method fusing oil chromatography characteristics and system thereof

Abstract

The invention relates to the technical field of transformer detection, in particular to a transformer fault evolution prediction method and system fusing oil chromatography characteristics, and provides the transformer fault evolution prediction method and system fusing the oil chromatography characteristics. The method comprises the following steps: firstly, acquiring transformer oil chromatographic data and equipment operation state data; secondly, respectively establishing neural network models to extract oil chromatography gas component features and equipment operation state features, and performing feature fusion through an attention mechanism; then, taking the fusion features as output variables for constraining a BP neural network model, and carrying out training learning; and finally, predicting the future fault probability of the transformer by using real-time monitoring data to obtain a fault evolution trend. According to the method, through an innovative algorithm and multi-source data fusion, the accuracy, timeliness and reliability of fault prediction are remarkably improved, and the method has excellent generalization ability and interpretability. The method has important practical value in the aspects of improving the safety of an electric power system, optimizing equipment maintenance, reducing operation and maintenance cost and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer faults, in particular to a method and system for predicting the evolution of transformer faults by integrating oil chromatogram characteristics. Background Art

[0002] As a key device in the power system, the operating state of a transformer directly affects the safety and stability of the entire power grid. With the rapid development of smart grids, higher requirements are put forward for the accuracy and timeliness of transformer fault prediction. Traditional transformer fault diagnosis methods mainly rely on regular oil chromatogram analysis and empirical judgment. Although certain achievements have been made in practical applications, there are still many limitations.

[0003] In recent years, with the progress of artificial intelligence technology, transformer fault prediction methods based on machine learning have gradually emerged. These methods analyze historical data to establish fault prediction models, which to a certain extent improve the accuracy and automation level of prediction. However, existing intelligent prediction methods still face some significant technical challenges.

[0004] Firstly, most existing methods only focus on a single data source, such as only using oil chromatogram data or only relying on equipment operating parameters. The limitation of this single data source results in the prediction model being unable to comprehensively grasp the operating state of the transformer, thus affecting the accuracy and reliability of prediction. Secondly, existing intelligent prediction algorithms often simplify transformer fault prediction into a purely data-driven problem, ignoring the valuable experience and knowledge accumulated by power domain experts. Although this approach can achieve good prediction results in some cases, it lacks interpretability and is difficult to gain the trust and adoption of actual operation and maintenance personnel.

[0005] In addition, existing methods still have deficiencies in dealing with the time-series characteristics of transformer operating data and the fusion of multi-source heterogeneous data. The evolution of transformer faults is often a complex time-series process involving multiple interrelated parameters. However, most existing methods fail to effectively capture this time-series dependence and are also difficult to achieve the organic fusion of different types of data, resulting in limitations in the timeliness and comprehensiveness of prediction results.

[0006] Finally, the generalization ability and robustness of existing intelligent prediction methods still need to be improved. Due to the diversity and complexity of transformer fault types and the dynamic changes in the actual operating environment, the prediction model needs to have strong adaptability and anti-interference ability. However, current methods often perform poorly when facing new types of faults or abnormal operating conditions and are difficult to meet the requirements of practical applications. Summary of the Invention

[0007] In view of the above problems, the present invention proposes a transformer fault evolution prediction method and system integrating oil chromatography characteristics. The method aims to achieve more accurate, reliable and forward-looking transformer fault prediction through the organic combination of multi-source data fusion, deep learning algorithms and expert knowledge.

[0008] The present invention proposes a transformer fault evolution prediction method and system integrating oil chromatography characteristics, including:

[0009] An acquisition step, including:

[0010] Acquire transformer oil chromatography data and equipment operation status data;

[0011] A processing step, including:

[0012] Based on the transformer oil chromatography data, establish a neural network model for transformer oil chromatography data, perform feature extraction of transformer oil chromatography data, and obtain oil chromatography gas component features;

[0013] Based on the equipment operation status data, establish a neural network model for transformer operation status data, perform feature extraction of equipment operation status, and obtain equipment operation status features;

[0014] Fuse the oil chromatography gas component features and equipment operation status features through an attention mechanism;

[0015] Use the features fused by the neural network model as the output variables of the constrained BP neural network model, and perform BP neural network training and learning based on the transformer sample data set;

[0016] An output step, including:

[0017] Use the concentration of oil chromatography gas components and equipment operation status parameters monitored in real time by the transformer as the input of the constrained BP neural network model for prediction, and output the probability of the transformer having a fault within a future set time interval, so as to obtain the fault evolution trend of the transformer.

[0018] Preferably, the specific implementation of establishing a neural network model for transformer oil chromatography data, using the transformer oil chromatography data as the input of the neural network model, performing feature extraction of transformer oil chromatography data, and obtaining oil chromatography gas component features is:

[0019] Use a convolutional neural network to achieve feature extraction of transformer oil chromatography data;

[0020] Extract 8 gases including H2, total hydrocarbons, CH4, C2H6, C2H2, C2H4, CO, and CO2 from the transformer oil chromatography data as the extraction objects;

[0021] Take the concentration value of each gas as the input of the convolutional neural network, obtain the component characteristics of the i-th (i = 1, 2, …, 8) gas, and establish a feature vector.

[0022] Take the time tag in the transformer oil chromatogram data and the batch number of the oil chromatogram test as additional constraint data, and combine the characteristics of each gas component in the transformer oil chromatogram to obtain the extraction result of the oil chromatogram data characteristics of 8 gases at the same time.

[0023] Preferably, the specific implementation of establishing the neural network model for transformer operation state data, taking the transformer operation state data as the input of the neural network model, and extracting the equipment operation state characteristics is as follows:

[0024] Use a convolutional neural network to implement the feature extraction of transformer operation state data.

[0025] Extract 5 feature elements, namely oil temperature rise, secondary voltage, load current, total load current, and reactive power ratio, from the transformer operation state data as the extraction objects.

[0026] Take the 5 kinds of data as the input of the convolutional neural network, obtain the corresponding features, establish the output variables of the convolutional neural network, and thus obtain the operation state characteristics of the equipment.

[0027] Preferably, the specific implementation of feature fusion of the oil chromatogram gas component characteristics and the equipment operation state characteristics through the attention mechanism is as follows:

[0028] According to the attention mechanism, normalize the extraction result of the oil chromatogram data characteristics and the operation state characteristics of the equipment respectively, splice them into a state feature map, and input it into the attention mechanism module for feature fusion.

[0029] Take the state feature map as the input variable of the attention mechanism module, and the result after global average pooling as the input variable of the Transformer module.

[0030] Use the Transformer module to implement feature fusion, fuse the transformer operation condition data and the oil chromatogram data, and obtain the fused features.

[0031] Preferably, the specific implementation of model training for constructing a constrained BP neural network for the transformer fault evolution prediction model based on the fused oil chromatogram features is as follows:

[0032] Take the fused features of the neural network model as the input variables of the constrained BP neural network model, and establish a constrained BP neural network model.

[0033] Based on the transformer sample data set, BP neural network training is carried out. The transformer oil chromatogram data, the transformer operation status, and the corresponding transformer fault probability form a training set and a validation set, which are respectively used to train the constrained BP neural network model;

[0034] For each sample in the sample set during the training process, the constrained BP neural network model is backpropagated to obtain a trained transformer fault evolution trend prediction model.

[0035] Preferably, the use of the trained transformer fault evolution trend prediction model for feature extraction and fault evolution trend prediction is specifically implemented as follows:

[0036] Use the trained transformer fault evolution trend prediction model to predict the feature extraction and fault evolution trend, and construct a data model based on the constrained feature extraction and time series fault prediction;

[0037] Input the transformer oil chromatogram data and the transformer operation status data as the input layer of the first type of data into the convolutional neural network of the constrained feature extraction model to complete the feature extraction of the data;

[0038] Calculate the cross entropy between the constrained feature extraction result of the transformer and the expert rule result, and use the cross entropy calculation result as a constraint term to constrain the current transformer oil chromatogram data;

[0039] Based on the constrained transformer oil chromatogram data, input it into the constrained BP neural network model to obtain the prediction result of the transformer oil chromatogram data;

[0040] Based on this result, perform temperature prediction based on the transformer operation status characteristics.

[0041] Preferably, the expert rule is specifically implemented as follows:

[0042] Combine the transformer operation condition data and the data features of the oil chromatogram data with the expert knowledge rule as a constraint term;

[0043] The expert knowledge rule includes: when the current state data of the transformer and the transformer oil chromatogram data do not conform to the expert knowledge rule, it is considered that the current state data and the transformer oil chromatogram data are abnormal.

[0044] Preferably, it further includes the following steps:

[0045] Define a prediction threshold constraint layer. According to the determination of the oil chromatogram fault level, use the threshold as a positive incentive and define the negative incentive as a preset value;

[0046] Add a positive incentive and a negative incentive to the loss function of the gas component content prediction part in the dual-stream neural network model to obtain an optimized loss function;

[0047] Among them, the minimum proportional coefficient of positive and negative incentives is used to control the proportion of positive and negative incentives.

[0048] Preferably, the following steps are further included:

[0049] Introduce expert rules into the neural network prediction framework, and establish a soft constraint term for transformer fault evolution prediction based on expert experience;

[0050] Use a method based on dynamic weights to add the soft constraint term to the neural network prediction framework;

[0051] Use the change of the transformer health state evaluation index in the operation state of the transformer equipment to add a soft constraint term to the neural network prediction model;

[0052] When the change rate of the transformer health index exceeds the preset threshold, it is considered that the health state of the transformer has deteriorated.

[0053] A transformer fault evolution prediction system integrating oil chromatogram characteristics includes:

[0054] A data acquisition module for acquiring transformer oil chromatogram data and equipment operation state data;

[0055] A neural network construction and feature extraction module for transformer oil chromatogram data, which is used to establish a neural network model for transformer oil chromatogram data, take the transformer oil chromatogram data as the input of the neural network model, and perform feature extraction of the transformer oil chromatogram data to obtain oil chromatogram gas component features;

[0056] A neural network construction and feature extraction module for transformer operation state data, which is used to establish a neural network model for transformer operation state data, take the transformer operation state data as the input of the neural network model, and perform feature extraction of the equipment operation state to obtain equipment operation state features;

[0057] A neural network model data fusion module for fusing the oil chromatogram gas component features and the equipment operation state features through an attention mechanism;

[0058] A constrained BP neural network model construction module for taking the features fused by the neural network model as the output variables of the constrained BP neural network model, and performing BP neural network training and learning based on the transformer sample data set;

[0059] A model actual application module for taking the concentration of oil chromatogram gas components and the equipment operation state parameters monitored in real time by the transformer as the input of the constrained BP neural network model for prediction, and outputting the probability of the transformer failing within a future set time interval, so as to obtain the fault evolution trend of the transformer.

[0060] The core of the present invention lies in solving key technical problems such as the limitations of a single data source, the lack of expert knowledge integration, insufficient processing of temporal characteristics, and poor model generalization ability in the prior art. By innovatively designing a dual-stream neural network architecture, the present invention realizes the effective integration of oil chromatogram data and equipment operation status data. The introduction of the attention mechanism and Transformer module effectively captures the temporal dependence relationship of the data. At the same time, by introducing prediction threshold constraints and expert rule soft constraints, domain expert knowledge is ingeniously incorporated into the deep learning model, improving the interpretability and credibility of the model.

[0061] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0062] Firstly, the present invention significantly improves the accuracy and timeliness of transformer fault prediction. Through multi-source data fusion and deep learning algorithms, this method can more comprehensively analyze the operation status of transformers and capture subtle fault signs. Experimental results show that the fault prediction accuracy of this method is as high as 95.8%, far exceeding traditional methods and single deep learning methods. More importantly, this method can issue a fault warning 15 days in advance, providing sufficient response time for maintenance personnel and greatly reducing the risk of sudden failures.

[0063] Secondly, the present invention effectively reduces the false alarm rate and improves the reliability of prediction results. By introducing expert rule soft constraints, this method controls the false alarm rate at a low level of 2.1%, which is of great significance for reducing unnecessary maintenance and downtime and lowering operation and maintenance costs. At the same time, this low false alarm rate also enhances the trust of maintenance personnel in the prediction results and is conducive to the popularization and application of the method in actual work.

[0064] Furthermore, the present invention demonstrates excellent model generalization ability. The average F1 score on different types of faults reaches 0.93, indicating that this method can effectively handle various types of transformer faults and adapt to complex and changing actual operating environments. This strong generalization ability stems from the unique multi-source data fusion strategy and deep learning architecture design of the present invention, enabling the model to learn more general fault characteristics.

[0065] Finally, the method of the present invention has good interpretability and operability. By incorporating expert knowledge into the model in the form of soft constraints, this method not only maintains high prediction performance but also retains the interpretability advantages of traditional methods. This design makes the prediction results easier to understand and accept by maintenance personnel and is conducive to the implementation and application of the method in actual work.

[0066] In summary, the transformer fault evolution prediction method and system integrating oil chromatogram features proposed by the present invention effectively solve the key problems existing in the prior art through innovative algorithm design and multi-source data fusion. This method not only significantly improves the accuracy, timeliness, and reliability of transformer fault prediction but also has excellent generalization ability and interpretability. These advantages make the present invention of great practical value in enhancing the safe and stable operation of power systems, optimizing equipment maintenance strategies, reducing operation and maintenance costs, etc., providing strong technical support for the development of smart grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is the overall logic block diagram of the system of the present invention;

[0068] Figure 2 is the flow chart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] Please refer to the attached Figure 1-2 , the present invention provides a transformer fault evolution prediction method and system integrating oil chromatogram features. This method realizes accurate prediction of transformer faults through deep learning technology and multi-source data fusion. The following is a detailed description of this method:

[0070] First, the method of the present invention includes an acquisition step, a processing step, and an output step. In the acquisition step, the system acquires transformer oil chromatogram data and equipment operation status data. These data are the basis for fault prediction. Among them, the oil chromatogram data reflects the content of gas components in transformer oil, while the equipment operation status data includes various operation parameters of the transformer.

[0071] In the processing step, the present invention adopts an innovative dual-stream neural network architecture. First, based on the transformer oil chromatogram data, a neural network model for transformer oil chromatogram data is established. This model extracts features from the oil chromatogram data through deep learning algorithms to obtain oil chromatogram gas component features. At the same time, based on the equipment operation status data, a neural network model for transformer operation status data is established to extract equipment operation status features.

[0072] Preferably, the present invention uses a convolutional neural network (CNN) to implement feature extraction. Taking the oil chromatogram data as an example, 8 gases including H2, total hydrocarbons, CH4, C2H6, C2H2, C2H4, CO, and CO2 are extracted as feature objects. The concentration value of each gas is used as the input of the CNN, and the corresponding feature vector is obtained through network processing. It should be noted that the present invention also innovatively introduces time tags and oil chromatogram test batch numbers as additional constraint data, which helps to improve the accuracy and temporal correlation of feature extraction.

[0073] For the equipment operation status data, the present invention selects five key parameters, namely, oil temperature rise, secondary voltage, load current, total load current, and reactive power ratio, as the objects for feature extraction. These parameters can comprehensively reflect the operation status of the transformer and are conducive to the accuracy of fault prediction.

[0074] The present invention adopts an innovative attention mechanism to fuse the characteristics of oil chromatogram gas components and the equipment operation status. Specifically, first, the two types of characteristics are normalized, and then spliced into a status feature map. This status feature map is used as the input of the attention mechanism module. After global average pooling, it is input into the Transformer module for feature fusion.

[0075] The use of the Transformer module is a major highlight of the present invention. It can effectively capture the long-range dependence relationships between different features, thereby achieving more accurate feature fusion. The fused features are used as the output variables to constrain the BP neural network model for subsequent fault prediction.

[0076] In the model training stage, the present invention conducts BP neural network training and learning based on the transformer sample data set. The training data includes transformer oil chromatogram data, transformer operation status, and the corresponding transformer fault probability. This multi-dimensional training data helps improve the generalization ability and prediction accuracy of the model.

[0077] In the output step, the present invention uses the concentration of oil chromatogram gas components and the equipment operation status parameters monitored in real time by the transformer as the input of the constrained BP neural network model for prediction. The model outputs the probability of the transformer having a fault within a set future time interval, thereby obtaining the fault evolution trend of the transformer.

[0078] A significant advantage of the method of the present invention is that it can simultaneously process oil chromatogram data and equipment operation status data, realizing the effective fusion of multi-source data. This fusion not only improves the prediction accuracy but also enhances the model's ability to identify different types of faults.

[0079] In addition, the attention mechanism and Transformer module introduced in the present invention enable the model to adaptively focus on important features, improving the effect of feature extraction and fusion. This is particularly important for fault prediction of complex systems such as transformers because different types of faults may be reflected in different features.

[0080] In practical applications, the method of the present invention can help power system operation and maintenance personnel timely discover potential transformer fault risks and take preventive measures in advance, thereby improving the reliability and safety of the power system. At the same time, since this method can predict the fault evolution trend, it can also provide a scientific basis for the maintenance and replacement decisions of transformers, optimize resource allocation, and reduce operation and maintenance costs.

[0081] Generally speaking, the transformer fault evolution prediction method integrating oil chromatogram features provided by the present invention realizes high-precision and high-reliability transformer fault prediction through innovative algorithm design and data processing methods, and has important theoretical value and practical application prospects.

[0082] In a preferred embodiment of the present invention, the oil chromatogram gas component features and the equipment operation state features are feature-fused through an attention mechanism. The specific implementation of this step reflects the innovation and technical depth of the present invention. First, the oil chromatogram data feature extraction results and the equipment operation state features are respectively normalized, and then spliced into a state feature map. This processing step helps to eliminate the dimensional differences between different features, making the subsequent fusion more reasonable.

[0083] Next, the state feature map is input into the attention mechanism module. The introduction of the attention mechanism is a major highlight of the present invention. It can adaptively focus on important features and improve the effect of feature fusion. Specifically, the attention mechanism module first performs global average pooling on the input state feature map, and the obtained result is used as the input variable of the Transformer module.

[0084] The use of the Transformer module is another important technical innovation of the present invention. In the complex task of transformer fault prediction, the Transformer can effectively capture the long-range dependence relationships between different features, thereby achieving more accurate feature fusion. Through the Transformer module, this method deeply fuses the transformer operation condition data and the oil chromatogram data to obtain the fused features. This fusion method not only improves the expression ability of the features, but also enhances the model's ability to identify different types of faults.

[0085] In terms of model training, the present invention uses a transformer fault evolution prediction model based on fused oil chromatogram features to construct a constrained BP neural network. The specific implementation of this step reflects the in-depth thinking of the present invention in model design and optimization. First, the features fused by the neural network model are used as the input variables of the constrained BP neural network model to establish a constrained BP neural network model. This design makes full use of the results of the previous feature fusion, ensuring the comprehensiveness and representativeness of the model input.

[0086] Preferably, the present invention trains a BP neural network based on a transformer sample dataset. The training data includes transformer oil chromatogram data, transformer operating status, and corresponding transformer fault probabilities. This multi-dimensional training data helps improve the model's generalization and prediction accuracy. During the training process, the constrained BP neural network model is backpropagated for each sample in the dataset, continuously optimizing the model parameters to ultimately produce a trained transformer fault evolution trend prediction model.

[0087] The method of the present invention also includes using the trained transformer fault evolution trend prediction model to perform feature extraction and fault evolution trend prediction. In this step, a data model based on constrained feature extraction and time series fault prediction is first constructed. This model design fully considers the temporal characteristics of transformer fault evolution, which helps improve prediction accuracy.

[0088] Specifically, this method uses transformer oil chromatogram data and transformer operating status data as the first-class data input layer and feeds them into a convolutional neural network in a constrained feature extraction model to extract features from the data. This step fully leverages the advantages of convolutional neural networks in feature extraction, effectively capturing local features and patterns in the data.

[0089] Next, a key innovation of this invention lies in its integration of expert knowledge. Specifically, a cross-entropy calculation is performed between the transformer constraint feature extraction results and the expert rule results. This result is used as a constraint term to constrain the current transformer oil chromatogram data. This approach cleverly combines machine learning with domain expert knowledge, ensuring the model's learning ability while incorporating the experience of human experts, effectively improving the reliability of predictions.

[0090] In a preferred embodiment of the present invention, the specific implementation of expert rules is also noteworthy. This method uses the combined data of transformer operating status data and oil chromatogram data features, along with expert knowledge rules, as constraints. These expert knowledge rules include the following: if the current transformer status data and oil chromatogram data do not conform to the expert knowledge rules, then the current status data and oil chromatogram data are considered abnormal. This rule setting fully considers practical engineering experience, effectively identifying and processing abnormal data, and improving the robustness of the model.

[0091] Finally, this method inputs the constrained transformer oil chromatogram data into a constrained BP neural network model to obtain prediction results. Based on these results, temperature prediction is further performed based on the transformer's operating status characteristics. This multi-step, multi-angle prediction method fully utilizes the characteristics of different data types and can comprehensively assess the transformer's operating status and potential risks.

[0092] Generally speaking, the transformer fault evolution prediction method integrating oil chromatogram features provided by the present invention realizes high-precision and high-reliability transformer fault prediction through innovative feature fusion, model training, and prediction strategies. This method not only has multiple innovations in algorithm design but also fully considers various factors in practical applications, possessing strong theoretical value and practicality.

[0093] In a preferred embodiment of the present invention, a prediction threshold constraint layer is introduced, which is an innovative design aimed at improving the accuracy and reliability of prediction. Specifically, according to the determination of the oil chromatogram fault level, this method takes the threshold as a positive incentive while defining the negative incentive as a preset value. This design fully considers the actual requirements of transformer fault diagnosis and can effectively distinguish different degrees of fault risks.

[0094] Preferably, the present invention optimizes the loss function of the gas component content prediction part in the dual-stream neural network model. By adding positive and negative incentives, the optimized loss function is obtained. A key innovation point of this optimization strategy is the introduction of the minimum ratio coefficient of positive and negative incentives to control the ratio of positive and negative incentives. For example, this ratio coefficient can be set to 0.2, which means the negative incentive is at least 20% of the positive incentive. The introduction of this parameter enables the model to better balance different types of prediction errors during the training process, thereby improving the overall prediction performance.

[0095] Another innovation point of the present invention is the introduction of expert rules into the neural network prediction framework. This approach reflects the effort of the present invention in integrating traditional knowledge and modern algorithms. Specifically, this method establishes soft constraint terms for transformer fault evolution prediction based on expert experience. These soft constraint terms are not hard rules but affect the prediction results of the model through a certain weight.

[0096] In practical applications, the present invention adopts a method based on dynamic weights to add soft constraint terms to the neural network prediction framework. The design of this dynamic weight takes into account the time-varying nature of the transformer operating state and can adaptively adjust the influence of constraint terms according to different operating stages. For example, a time window of 30 days can be set, and within this window, the weights of the constraint terms are dynamically adjusted, enabling the model to better adapt to short-term state changes.

[0097] In addition, the present invention innovatively utilizes the variation of the transformer health status evaluation index in the equipment operation state and adds a soft constraint term to the neural network prediction model. A key point of this design lies in setting a threshold for the change rate of the transformer health index. When the change rate of the health index exceeds the preset threshold, the model will consider that the health status of the transformer has deteriorated. For example, this threshold can be set to 5%, that is, when the change rate of the health index within a certain period of time (such as 7 days) exceeds 5%, a warning of deteriorated health status is triggered.

[0098] Finally, the present invention also provides a complete transformer fault evolution prediction system integrating oil chromatogram characteristics. The system includes multiple functional modules, and each module has its specific responsibilities and functions.

[0099] The data acquisition module 1 is responsible for acquiring transformer oil chromatogram data and equipment operation state data. This module can collect data through various methods, such as real-time monitoring systems, regular sampling and analysis, etc., to ensure the comprehensiveness and timeliness of the data.

[0100] The neural network construction and feature extraction module 2 for transformer oil chromatogram data is used to establish a neural network model for transformer oil chromatogram data. This module takes the transformer oil chromatogram data as the input of the neural network model, conducts feature extraction of the transformer oil chromatogram data, and finally obtains the characteristics of the oil chromatogram gas components. The design of this module fully considers the characteristics of the oil chromatogram data and can effectively extract key gas component information.

[0101] The neural network construction and feature extraction module 3 for transformer operation state data is responsible for establishing a neural network model for transformer operation state data. This module takes the transformer operation state data as the input of the neural network model, conducts feature extraction of the equipment operation state, and obtains the equipment operation state characteristics. The design of this module considers multiple key parameters of the transformer operation and can comprehensively reflect the operation state of the equipment.

[0102] The neural network model data fusion module 4 is one of the core modules of this system. It fuses the characteristics of the oil chromatogram gas components and the equipment operation state characteristics through an attention mechanism. This fusion method can make full use of the advantages of different types of data and improve the accuracy of prediction.

[0103] The constrained BP neural network model construction module 5 takes the fused characteristics of the neural network model as the output variable of the constrained BP neural network model and conducts BP neural network training and learning based on the transformer sample data set. The design of this module reflects the innovation of the present invention in model optimization.

[0104] Finally, the model actual application module 6 is responsible for using the concentration of oil chromatographic gas components and the equipment operation status parameters real-time monitored by the transformer as the input to constrain the BP neural network model for prediction, and outputting the probability of the transformer failing within the set time interval in the future, so as to obtain the fault evolution trend of the transformer.

[0105] Generally speaking, the transformer fault evolution prediction method and system integrating oil chromatographic characteristics provided by the present invention achieve high-precision and high-reliability transformer fault prediction through innovative algorithm design, multi-source data fusion, and expert knowledge integration. This method and system are not only innovative in theory, but also have strong operability and practical value in actual applications, and are expected to provide important support for the safe and stable operation of power systems.

[0106] In order to verify the effectiveness and superiority of the transformer fault evolution prediction method and system integrating oil chromatographic characteristics proposed by the present invention, a series of simulation experiments were designed. These experiments were based on real transformer operation data and simulated different operating conditions and fault scenarios.

[0107] The experimental settings are as follows:

[0108] The main transformer of a 500 kV substation was selected as the research object, and its operation data for nearly 5 years were collected, including oil chromatographic data and equipment operation status data. The data set included data during normal operation and data of several known fault events. 80% of the data was used for model training, and 20% was used for testing.

[0109] Example 1 (the method of the present invention):

[0110] The transformer fault evolution prediction method integrating oil chromatographic characteristics proposed by the present invention was adopted, including innovative points such as a dual-stream neural network architecture, attention mechanism feature fusion, prediction threshold constraint, and expert rule soft constraint.

[0111] Comparative Example 1:

[0112] Only the traditional oil chromatographic analysis method was used for fault prediction, mainly based on gas content and ratio rules.

[0113] Comparative Example 2:

[0114] A single deep learning model (such as LSTM) was used for fault prediction, and only oil chromatographic data was used as the input.

[0115] The following key indicators were selected to evaluate the performance of the method:

[0116] 1. Fault prediction accuracy rate: the proportion of correctly predicted faults.

[0117] 2. Fault warning lead time: How long in advance can a warning be given before an actual fault occurs?

[0118] 3. False alarm rate: The proportion of faults that are predicted incorrectly.

[0119] 4. F1 score: An indicator that comprehensively considers precision and recall.

[0120] 5. Model generalization ability: The prediction performance on different types of faults.

[0121] The experimental results are shown in the following table:

[0122]

[0123]

[0124] The model generalization ability is measured by the average F1 score on different types of faults.

[0125] As can be seen from the experimental results, the method of the present invention (Example 1) performs excellently in all indicators, significantly better than the traditional method (Comparative Example 1) and the single deep learning method (Comparative Example 2). The specific analysis is as follows:

[0126] 1. Fault prediction accuracy: The method of the present invention achieves a high accuracy of 95.8%, far higher than the other two methods. This is mainly due to the dual-stream neural network architecture of the present invention, which can simultaneously process oil chromatogram data and equipment operation status data, providing a more comprehensive basis for fault prediction.

[0127] 2. Fault warning lead time: The method of the present invention can give a warning about potential faults 15 days in advance, 10 days earlier than the traditional method and 5 days earlier than the single deep learning method. This significant early warning ability stems from the attention mechanism and prediction threshold constraint introduced in the present invention, enabling the model to capture fault signs earlier.

[0128] 3. False alarm rate: The false alarm rate of the method of the present invention is only 2.1%, far lower than the other two methods. This shows that the expert rule soft constraint introduced in the present invention effectively improves the reliability of prediction and reduces the occurrence of false alarms.

[0129] 4. F1 score: The F1 score of the method of the present invention reaches 0.967, indicating that it achieves a good balance between precision and recall. This balance is very important for reducing missed alarms and false alarms in practical applications.

[0130] 5. Model generalization ability: The average F1 score of the method of the present invention on different types of faults reaches 0.93, showing excellent generalization ability. This is mainly due to the multi-source data fusion and deep learning architecture of the present invention, enabling the model to learn more general fault features.

[0131] It is worth noting that in terms of the indicator of the fault warning lead time, the method of the present invention performs particularly outstandingly. It can give a warning of potential faults 15 days in advance, which is of great significance for the safe operation of the power system and the formulation of maintenance plans. It provides sufficient time for the operation and maintenance personnel to formulate coping strategies, greatly reducing the risks and losses brought by sudden faults.

[0132] In addition, the excellent performance of the method of the present invention in terms of the false alarm rate is also worthy of attention. In practical applications, too high a false alarm rate will lead to unnecessary repairs and shutdowns, increasing the operation and maintenance costs. The method of the present invention controls the false alarm rate at a low level of 2.1%, which means that it can provide a more reliable basis for operation and maintenance decisions.

[0133] In summary, the transformer fault evolution prediction method integrating oil chromatography characteristics proposed by the present invention shows obvious advantages in various key indicators. It not only improves the accuracy and timeliness of fault prediction, but also reduces the false alarm rate and enhances the generalization ability of the model. These advantages fully prove the innovation and practical value of the present invention in the field of transformer fault prediction, and are expected to provide strong technical support for the safe and stable operation of the power system.

[0134] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A transformer fault evolution prediction method integrating oil chromatogram characteristics, characterized in that Including: An acquisition step, including: Acquiring transformer oil chromatogram data and equipment operation status data; A processing step, including: Based on the transformer oil chromatogram data, establishing a neural network model for transformer oil chromatogram data, performing feature extraction on the transformer oil chromatogram data, and obtaining oil chromatogram gas component features; Based on the equipment operation status data, establishing a neural network model for transformer operation status data, performing feature extraction on the equipment operation status, and obtaining equipment operation status features; Performing feature fusion on the oil chromatogram gas component features and the equipment operation status features through an attention mechanism; Using the features after fusion by the neural network model as the output variables of a constrained BP neural network model, and performing BP neural network training and learning based on a transformer sample data set; An output step, including: Taking the concentration of oil chromatogram gas components and the equipment operation status parameters monitored in real time for the transformer as the input of the constrained BP neural network model for prediction, and outputting the probability of the transformer failing within a set future time interval, thereby obtaining the fault evolution trend of the transformer.

2. The transformer fault evolution prediction method integrating oil chromatography characteristics according to claim 1, wherein The specific implementation of establishing the neural network model for transformer oil chromatogram data, using the transformer oil chromatogram data as the input of the neural network model, and performing feature extraction on the transformer oil chromatogram data to obtain oil chromatogram gas component features is as follows: Using a convolutional neural network to implement feature extraction of transformer oil chromatogram data; Extracting 8 gases, namely H2, total hydrocarbons, CH4, C2H6, C2H2, C2H4, CO, and CO2, from the transformer oil chromatogram data as extraction objects; Taking the concentration value of each gas as the input of the convolutional neural network, obtaining the component features of the corresponding i-th (i = 1, 2,..., 8) gas, and establishing a feature vector; Taking the time label in the transformer oil chromatogram data and the batch number of the oil chromatogram test as additional constraint data, and combining the oil chromatogram gas component features of each gas to obtain the oil chromatogram data feature extraction result of 8 gases at the same time.

3. The transformer fault evolution prediction method integrating oil chromatogram features according to claim 1, wherein, The specific implementation of establishing the neural network model for transformer operation status data, using the transformer operation status data as the input of the neural network model, and performing feature extraction on the equipment operation status to obtain equipment operation status features is as follows: Using a convolutional neural network to implement feature extraction of transformer operation status data; Extracting 5 feature elements, namely oil temperature rise, secondary voltage, load current, total load current, and reactive power ratio, from the transformer operation status data as extraction objects; Taking the 5 types of data as the input of the convolutional neural network, obtaining the corresponding features, and establishing the output variables of the convolutional neural network, thereby obtaining the equipment operation status features.

4. The transformer fault evolution prediction method integrating oil chromatogram features according to claim 1, characterized in that The specific implementation of performing feature fusion on the oil chromatogram gas component features and the equipment operation status features through an attention mechanism is as follows: According to the attention mechanism, normalizing the oil chromatogram data feature extraction result and the equipment operation status features respectively, splicing them into a state feature map, and inputting it into the attention mechanism module for feature fusion; Taking the state feature map as the input variable of the attention mechanism module, and using the result after global average pooling as the input variable of the Transformer module; The Transformer module is adopted to achieve feature fusion, fusing the transformer operation condition data and the oil chromatogram data to obtain the fused features.

5. The transformer fault evolution prediction method integrating oil chromatography characteristics according to claim 1, characterized in that The transformer fault evolution prediction model based on the fused oil chromatogram features constructs a constrained BP neural network. The specific implementation of model training is as follows: Take the features fused by the neural network model as the input variables of the constrained BP neural network model to establish a constrained BP neural network model; Based on the transformer sample data set, conduct BP neural network training. The transformer oil chromatogram data, the transformer operation state and the corresponding transformer fault probability form a training set and a verification set, which are respectively used to train the constrained BP neural network model; For each sample in the sample set during the training process, perform backpropagation on the constrained BP neural network model to obtain the trained transformer fault evolution trend prediction model.

6. The transformer fault evolution prediction method integrating oil chromatogram characteristics according to claim 1, wherein, The use of the trained transformer fault evolution trend prediction model for feature extraction and fault evolution trend prediction is specifically implemented as follows: Use the trained transformer fault evolution trend prediction model to predict the feature extraction and the fault evolution trend, and construct a data model based on the constrained feature extraction and time series fault prediction; Take the transformer oil chromatogram data and the transformer operation state data as the input layer of the first type of data and input them into the convolutional neural network of the constrained feature extraction model to complete the feature extraction of the data; Calculate the cross entropy between the constrained feature extraction result of the transformer and the expert rule result, and use the cross entropy calculation result as a constraint term to constrain the current transformer oil chromatogram data; Based on the constrained transformer oil chromatogram data, input it into the constrained BP neural network model to obtain the prediction result of the transformer oil chromatogram data; Based on this result, conduct temperature prediction based on the transformer operation state characteristics.

7. The transformer fault evolution prediction method integrating oil chromatogram features according to claim 2, characterized in that The specific implementation of the expert rule is as follows: Take the data obtained by combining the transformer operation condition data and the oil chromatogram data features and the expert knowledge rule as the constraint term; The expert knowledge rule includes: when the current state data of the transformer and the transformer oil chromatogram data do not conform to the expert knowledge rule, it is considered that the current state data and the transformer oil chromatogram data are abnormal.

8. The method for predicting the evolution of transformer faults by fusing oil chromatographic characteristics according to claim 1, characterized in that It also includes the following steps: Define a prediction threshold constraint layer. According to the determination of the oil chromatogram fault level, take the threshold as a positive incentive and define the negative incentive as a preset value; Add positive and negative incentives to the loss function of the gas component content prediction part in the dual-stream neural network model to obtain an optimized loss function; Among them, the minimum proportional coefficient of the positive and negative incentives is used to control the ratio of the positive and negative incentives.

9. The transformer fault evolution prediction method integrating oil chromatography characteristics according to claim 1, wherein, It also includes the following steps: Introduce the expert rule into the neural network prediction framework and establish a soft constraint term for transformer fault evolution prediction based on expert experience; Use the method based on dynamic weights to add the soft constraint term to the neural network prediction framework; Use the change of the transformer health state evaluation index in the transformer equipment operation state to add a soft constraint term to the neural network prediction model; When the change rate of the transformer health index exceeds the preset threshold, it is considered that the transformer health state has deteriorated.

10. A transformer fault evolution prediction system for fusing oil chromatography features for performing the method according to any one of claims 1-9, characterized in that, Include: A data acquisition module for acquiring transformer oil chromatogram data and equipment operation status data; A neural network construction and feature extraction module for transformer oil chromatogram data, which is used to establish a neural network model for transformer oil chromatogram data, take the transformer oil chromatogram data as the input of the neural network model, perform feature extraction on the transformer oil chromatogram data, and obtain the characteristics of oil chromatogram gas components; A neural network construction and feature extraction module for transformer operation status data, which is used to establish a neural network model for transformer operation status data, take the transformer operation status data as the input of the neural network model, perform feature extraction on the equipment operation status, and obtain the equipment operation status characteristics; A neural network model data fusion module for fusing the characteristics of the oil chromatogram gas components and the equipment operation status characteristics through an attention mechanism; A constrained BP neural network model construction module for taking the features after fusion of the neural network model as the output variables of the constrained BP neural network model and performing BP neural network training and learning based on the transformer sample data set; A model practical application module for predicting by taking the concentration of oil chromatogram gas components and the equipment operation status parameters monitored in real time by the transformer as the input of the constrained BP neural network model, and outputting the probability of the transformer failing within a set future time interval, so as to obtain the fault evolution trend of the transformer.

Images