Intelligent evaluation method for oil-immersed transformer state
The transformer condition assessment method optimized by the non-coded ratio method and the IPSO-XGBoost algorithm solves the problems of long training time and low accuracy in the existing technology, and achieves efficient and accurate transformer condition assessment, thereby improving the safety and stability of the power system.
Patent Information
- Application Number
- CN202310115667.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-15
- Publication Date
- 2025-11-28
- Estimated Expiration
- 2043-02-15
AI Technical Summary
Existing machine learning models suffer from long training times, low intelligent assessment accuracy, and low learning efficiency in transformer condition assessment, making it difficult to achieve efficient and accurate transformer condition monitoring and assessment.
The DGA data is processed using the non-coded ratio method, and a transformer evaluation model based on XGBoost is established. The key hyperparameters are optimized by improving the particle swarm optimization algorithm (IPSO-XGBoost) to improve the intelligent evaluation performance of the model.
This improved the accuracy and stability of transformer condition assessment, ensured the reliability of assessment results, shortened training time, enhanced the real-time detection capability of transformer condition, and safeguarded the safety and stability of the power system.
Smart Images
Figure CN116186010B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The intelligent evaluation method for the state of an oil-immersed transformer belongs to the technical field of electric power and particularly relates to an intelligent evaluation method for the state of an oil-immersed transformer. BACKGROUND
[0002] At present, there are hidden troubles such as long operation life and insulation deterioration in the power transformers in operation of the power system. An analysis method capable of monitoring and evaluating the state of the transformer in real time needs to be proposed, and an efficient and accurate evaluation model needs to be established to accurately predict the operation state of the transformer and effectively cope with the hidden troubles of the transformer, thereby ensuring the safety, stability and reliability of the operation of the power system.
[0003] With the occurrence of mechanical faults such as electrical aging and thermal aging, a variety of gases are dissolved in the transformer oil during the operation of the transformer. Therefore, the dissolved gas analysis (DGA) of the transformer oil can timely find the potential fault risk of the transformer, and is a recognized feasible method for evaluating the state of the transformer in the electric power industry. With the emergence of artificial intelligence and deep learning, the intelligent evaluation method for the state of the transformer based on machine learning has become a topic of interest in the research field, and the transformer state evaluation model based on artificial intelligence has to a great extent replaced the traditional state evaluation method and become a mainstream tool for detecting the operation state of the transformer. At present, among the machine learning models excavated based on the DGA data as characteristic parameters, there are artificial neural networks, support vector machines, correlation vector machines and hybrid models with optimized parameters. However, the above machine learning models have problems such as long training time, insufficient intelligent evaluation accuracy, low learning efficiency and the like.
[0004] In view of this, it is necessary to study an intelligent evaluation method for the state of an oil-immersed transformer, which introduces an algorithm to optimize the key parameters of the machine learning model, further improves the accuracy and reliability of the intelligent evaluation of the state of the transformer, and makes up for and solves the above problems. SUMMARY
[0005] The intelligent evaluation method for the state of an oil-immersed transformer overcomes the deficiencies of the prior art and solves the technical problem of providing an intelligent evaluation method for the state of an oil-immersed transformer, which exhibits stronger stability and higher evaluation accuracy in analyzing the DGA sample data of the transformer. When evaluating the state of the transformer, the intelligent evaluation method can efficiently train the model while ensuring high accuracy of evaluation and high reliability of the result.
[0006] To solve the above technical problems, the technical scheme adopted by the present application is as follows: an intelligent evaluation method for the state of an oil-immersed transformer, comprising the following steps:
[0007] S1. Collect data from online transformer monitoring and oiling experiments, statistically analyze the composition and content of gases in transformer oil, and establish a DGA sample dataset.
[0008] S2. Perform data filtering on the DGA sample data to remove invalid and interfering data;
[0009] S3. The DGA data is preprocessed using the uncoded ratio method to characterize, normalize, and standardize the sample data;
[0010] S4. Determine the feature dimension and encode the corresponding transformer state as the evaluation result of the model output;
[0011] S5. Balance the data samples, group them, and randomly select 80% of the sample data as the training set, and the rest as the test set;
[0012] S6. Establish a transformer evaluation model based on XGBoost;
[0013] S7. An IPSO-XGBoost evaluation model is proposed by using an improved particle swarm optimization to optimize the key hyperparameters of XGBoost.
[0014] S8. Import the training set into the IPSO-XGBoost evaluation model for machine training, and import the test set into the IPSO-XGBoost evaluation model to perform intelligent evaluation of the transformer state.
[0015] Beneficial Effects: This invention provides an intelligent assessment method for the condition of oil-immersed transformers. Based on the XGBoost classification model of machine learning, it uses a non-coded ratio method to process DGA data and employs the proportional relationships of nine feature dimensions as feature parameters for machine mining. Using the objective function and key hyperparameters of the classification model as optimization objectives, the IPSO algorithm is introduced to optimize the XGBoost classification model, improving its intelligent assessment performance and enhancing the reliability of the assessment results. A case study analysis of 1614 data points was conducted in the model. The IPSO-XGBoost intelligent assessment model established by this invention effectively solves problems such as slow particle swarm optimization speed and getting trapped in local optima, improving the accuracy and stability of transformer condition assessment. This method has certain guiding significance for the real-time and effective detection of transformer condition and ensuring the safety, stability, and reliability of power system operation. Attached Figure Description
[0016] The present invention will now be described in further detail with reference to the accompanying drawings;
[0017] Figure 1 This is a flowchart illustrating an intelligent assessment method for the condition of an oil-immersed transformer according to the present invention.
[0018] Figure 2 A specific flowchart of intelligent evaluation of the transformer state in step S8 of the present application is shown in the figure;
[0019] Figure 3 A schematic diagram of XGBoost decision tree inference state in the present application is shown in the figure;
[0020] Figure 4 A transformer state distribution graph of sample data in the embodiment of the present application is shown in the figure;
[0021] Figure 5 A schematic diagram of the optimization results comparison between the traditional PSO and IPSO in the embodiment of the present application is shown in the figure;
[0022] Figure 6 A process schematic diagram of the learning rate of the IPSO optimization model parameter in the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION
[0023] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the present application will be described clearly and completely below in combination with the embodiments in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0024] As shown in the figure, the present application provides an intelligent evaluation method of oil-immersed transformer state, which comprises the following steps: Figure 1
[0025] S1, transformer online monitoring and oil analysis data collection, statistics of the composition of the gas in the transformer oil and the content of each gas, and establishment of DGA sample data set;
[0026] S2, data screening is performed on the DGA sample data, and invalid data and interference data are cleaned;
[0027] S3, the DGA data is preprocessed by using the non-coding ratio method, so as to characterize, normalize and standardize the sample data;
[0028] S4, the feature quantity dimension is determined, and the corresponding transformer state is coded as the evaluation result of the model output;
[0029] S5, the data sample is balanced, grouped, and 80% of the sample data is randomly extracted as the training set, and the rest is taken as the test set;
[0030] S6, a transformer evaluation model based on XGBoost is established;
[0031] S7, adopt improved particle swarm optimization XGBoost key hyperparameters, propose IPSO-XGBoost evaluation model;
[0032] S8, import training set in IPSO-XGBoost evaluation model for machine training, import test set in IPSO-XGBoost evaluation model, and intelligently evaluate the state of the transformer.
[0033] In step S1, the composition of the gas in the transformer oil includes H2, CH4, C2H2, C2H4, C2H6 and various alkane and unsaturated hydrocarbon gases.
[0034] In step S2, the DGA sample data set is subjected to data screening, and the invalid data and interference data cleaned include but are not limited to data with less than 3 kinds of composition of the gas in the transformer oil and data with a difference of more than 3 orders of magnitude between the statistical gas content and the historical average data.
[0035] In step S3, the DGA data is preprocessed by using the following 9 kinds of gas content ratio relationship to form characteristic quantities:
[0036] CH4 / H2, C2H4 / C2H2, C2H4 / C2H6, (CH4+C2H4) / (C1+C2), H2 / (H2+C1+C2), C2H4 / (C1+C2), CH4 / (C1+C2), C2H6 / (C1+C2), C2H2 / (C1+C2).
[0037] Wherein, C1 is the sum of the content of alkane gas represented by CH4; C2 is the sum of the content of unsaturated hydrocarbon gas represented by C2H6, C2H4 and C2H2.
[0038] In step S4, the specific form of encoding the state of the transformer is:
[0039] According to the DL / T 722-2014 guide, the transformer state output corresponds to 8 state encodings:
[0040] Normal state is 0; Arc discharge is 1; Low energy discharge is 2; Partial discharge is 3; High temperature overheating is 4; Medium temperature overheating is 5; Low temperature overheating is 6; High energy discharge is 7.
[0041] In step S5, the method of grouping the balanced data samples is: classify and count the DGA data of different transformer states, and then proportionally and evenly distribute them to 5 groups to ensure the balance of data between groups.
[0042] The process of establishing the transformer evaluation model based on XGBoost in step S6 is:
[0043] S61, constructing a transformer evaluation model of XGBoost
[0044]
[0045] F = {f(x) = ω q (x)} q:R m →T,ω∈R T (2)
[0046] In formula (1), represents the final prediction result when training to the Kth classification and regression tree (CART), and each f k function corresponds to a CART; x i represents the ith sample; in formula (2), q and ω represent the structure vector and leaf weight of CART; R represents the space of the regression tree; m represents the number of features; T represents the number of leaf nodes; F is a set composed of all CARTs; x represents a certain sample; φ represents the functional relationship between x i and ; ω q represents the prediction value of the sample classification of the CART classifier;
[0047] After the sample is input, the structure vector q is classified according to the CART classification to the corresponding leaf label, each leaf label has different weights, and ω q is the prediction value of the sample classification of the CART classifier;
[0048] All CART classifiers in the set independently classify the input sample into leaf labels, and finally the XGBoost model accumulates the prediction values of all CART classifiers to obtain the prediction result;
[0049] S62, establishing a target function
[0050] After inputting the sample i, the regularization target function for the kth iteration is:
[0051]
[0052]
[0053] In the formula, L represents the original loss function; l represents the cumulative statistics of the prediction result; represents the prediction result of the ith sample; y i represents the prediction value of the previous round of model; Ω represents the regularization term, γ and λ are constants for constraining the regularization degree, and T represents the number of leaf nodes and ω 2 represents the leaf weight;
[0054] S63, optimization of the target function
[0055] The prediction value at the kth iteration is equal to the prediction value at the (k-1)th iteration plus f k The objective function is further written as:
[0056]
[0057] The objective function is expanded by the second-order Taylor expansion:
[0058]
[0059] In the formula, L (k) represents the objective function using the second-order Taylor expansion; g i and h i are the first-order and second-order derivatives of the loss function, respectively;
[0060] The constant part cannot be optimized, and the function part that can be optimized is:
[0061]
[0062] In the formula, ω represents the optimized objective function after ignoring the constant part;
[0063] Each sample will eventually be classified to a leaf node, and the samples in the same leaf node are recombined. The loss function of each sample is added up, and the process is as follows (assuming that the leaf node is numbered j):
[0064]
[0065] In formula (8), ω j represents the weight of the jth leaf node;
[0066] I j is the number set of all samples:
[0067] I j ={i|q(x i )=j} (9)
[0068] The objective function is rewritten as a monomial quadratic function about the leaf node prediction value ω, so the vertex formula is directly used to solve the optimal ω and the objective function value:
[0069]
[0070]
[0071] In the formula, ω is the optimal solution of ω j ;
[0072] Step S64, introducing a split node algorithm
[0073] In the training of XGBoost, a greedy algorithm is introduced to find the best split point more simply and quickly. When the amount of sample data increases, the greedy algorithm will consume a lot of test time to traverse all possible split points, at this time, an approximate algorithm is introduced to replace the greedy algorithm.
[0074] The step S7 adopts the improved particle swarm optimization XGBoost key hyperparameter, and the specific process of the IPSO-XGBoost evaluation model is as follows:
[0075] S71, particle swarm optimization algorithm design
[0076] In the particle swarm algorithm, the optimization direction of particles (potential solutions) in the D-dimensional target search space (solution space) is affected by three factors: the speed of the last iteration, individual cognition and group guidance. Individual cognition is the adjustment of the optimization direction of the next iteration by the particle according to its own learning; group guidance is the adjustment of the optimization direction of the next iteration by the particle according to the optimal action route of the group. Therefore, the inertia weight, acceleration factor and other parameters are introduced to describe the speed update and position update of the i-th particle:
[0077] V ij (k+1)=ωV ij (k)+c1r1(k)[p ij (k)-x ij (k)]+c2r2(k)[p gj (k)-x ij (k)] j∈D (12)
[0078] x ij (k+1)=x ij (k)+V ij (k+1) j∈D (13)
[0079] In the formula, V ij (k) represents the j-dimensional speed component of the i-th particle at the k-th iteration; ω is the inertia weight of the particle iteration; c1, c2 are the acceleration factors of the particle iteration; r1, r2 represent random numbers between 0 and 1; p ij (k) represents the j-dimensional component of the local optimal position of the i-th particle at the k-th iteration; x ij (k) represents the j-dimensional position component of the i-th particle at the k-th iteration; p gj (k) represents the j-dimensional component of the overall optimal position of the i-th particle at the k-th iteration;
[0080] S72, improved particle swarm optimization algorithm
[0081] For the three parameters in formula (12), (13), ω describes the "inertia" of the particle, which should be larger at the early stage of iteration to ensure that the particle independently flies to search the space, and should be smaller at the late stage of iteration to learn from other particles; c1 and c2 respectively describe the influence of the local optimal position and the overall optimal position, c1 should be larger at the early stage, and c2 should be larger at the late stage, so as to balance the global search ability and the local search ability of the particle. In view of this, the traditional particle swarm algorithm is improved, and the improved particle swarm (IPSO) algorithm is proposed, and the above parameters are improved by using the following formula:
[0082] ω(k)=ω max -(ω max -ω min )(k / T max ) 2 (14)
[0083] c1(k)=c 10 -(c 10 -c 11 )(k / T max ) (15)
[0084] c2(k)=c 20 -(c 20 -c 22 )(k / T max ) (16)
[0085] In the formula, ω(k), c1(k) and c2(k) represent the inertia weight and acceleration factor of the particle at the kth iteration, which are used to replace ω, c1 and c2 in formula (12); ω max , ω min respectively represent the maximum and minimum values of the inertia weight; c 10 , c 11 and c 20 , c 22 respectively represent the initial value and the final value of the acceleration factor c1 and c2; T max represents the maximum number of iterations;
[0086] S73, establish IPSO-XGBoost evaluation model
[0087] The IPSO algorithm is introduced into the XGBoost transformer evaluation model established in step S6, the global search and local search abilities of the particle optimization are balanced, the multiple key hyperparameters of XGBoost are simultaneously optimized by IPSO, and the IPSO-XGBoost evaluation model is established.
[0088] The key hyperparameters in step S73 include, but are not limited to, the number of classifiers, the minimum sum of sample weights in the child nodes, the learning rate, the depth of the tree, and the seed for the random generation function.
[0089] like Figure 2 As shown, the specific process of step S8, which involves importing the training set into the IPSO-XGBoost evaluation model for machine training and importing the test set into the IPSO-XGBoost evaluation model to intelligently evaluate the transformer state, is as follows:
[0090] S81. Import the training set into the IPSO-XGBoost evaluation model and perform machine training;
[0091] S82 and IPSO calculate the corresponding particle fitness for model hyperparameters;
[0092] S83 and IPSO search the population space, update the best position of particles and the globally optimal particle, and further update the particle search speed, position and inertia weight;
[0093] S84. Determine whether the model's evaluation result matches the actual state of the transformer. If not, repeat steps S82 and S83. If yes, output the optimal parameter particle to achieve automatic optimization of the model's hyperparameters.
[0094] S85. Input the test set into the evaluation model after the hyperparameters have been optimized, perform transformer state evaluation, and output the model evaluation results.
[0095] S86. Analyze the performance of the evaluation model based on the evaluation results.
[0096] Figure 3 This is a schematic diagram of the XGBoost decision tree inference state. For ease of understanding, the classification process of the "medium-temperature overheating" state type decision tree for transformers is used as an example. First, the DGA data, after being processed without encoding ratios, is input into the CART classifier; second, the CART classifier classifies the data based on the tree structure constructed by machine learning; finally, the classification results are assigned to CART leaf labels, each with a different weight. Different weights mean that the CART classifier will give different prediction values for different DGA data. Taking a leaf label with a weight of 0.607407451 as an example, the DGA data satisfies the following conditions:
[0097] 1. C2H6 / (C1+C2) < 0.695413113;
[0098] 2. C2H4 / C2H6 < 0.00537877344;
[0099] 3. CH4 / H2 ≥ 0.00028502199;
[0100] 4. H2 / (H2+C1+C2)≥0.422586203.
[0101] This indicates that the weight of the transformer state type of medium temperature overheating is 0.607407451. The data meeting other conditions correspond to different leaf weights, and the greater the weight, the greater the possibility of the transformer state type being medium temperature overheating.
[0102] Embodiment
[0103] To better reflect the evaluation effect of the intelligent evaluation method for the state of the oil-immersed transformer proposed in the present application, the following will be described through specific example analysis. The example analysis data is derived from literature, IEC TC 10 database, and part of the transformer online monitoring and oil chemical experimental data of the Southern Power Grid.
[0104] There are 3447 groups of transformer state statistical data, and after data cleaning, invalid data and interference data in the data sample are processed, 3228 groups of valid data of transformer state statistics are obtained. In order to ensure the balance of the data sample and the objectivity of the experimental results, the DGA data of different transformer states are classified and counted from the 3228 valid data, and then are evenly distributed in the 5 groups in proportion, so as to ensure the balance of the data between the groups. Each group is composed of 1614 groups of transformer DGA data determined by the state, and the ratio of the training set and the test set is 8:2, that is, the training set is 1291 groups and the test set is 323 groups. The transformer state distribution of the sample data of the embodiment of the present application is shown in Figure 4 . The detailed distribution of the training set of each group of data is shown in Table 1.
[0105] Table 1 Sample distribution
[0106]
[0107] Further, as shown in Figure 2 , the training set is imported into the IPSO-XGBoost evaluation model for machine training, and the test set is imported into the IPSO-XGBoost evaluation model for intelligent evaluation of the transformer state.
[0108] It is supplemented here that the intelligent evaluation model of the present application is in the environment of Inter(R) Core(TM) i7-10700, 2.90GHz, memory 16GB, uses Pycharm (Professional Edition) as the experimental simulation platform and Python 3.9 for programming.
[0109] The pre-processed DGA data is divided into non-coding ratios as feature parameters input into XGBoost, and 5 groups of DGA data are trained under default parameters, and the detailed evaluation results of the 5 test sets are shown in Table 2. The test results show that XGBoost has a high accuracy rate for evaluating the five states of the transformer: normal state, high-energy discharge, high-temperature overheating, medium-temperature overheating, and arc discharge, among which the evaluation accuracy of the normal state is the highest, and the average accuracy of the 5 test set evaluation is 0.940; while XGBoost has a lower accuracy rate for evaluating the three states of the transformer: low-energy discharge, partial discharge, and low-temperature overheating, among which the evaluation accuracy of low-energy discharge is the lowest, and the average accuracy of the 5 test set evaluation is 0.779. As shown in Table 1, the training samples of these three types of transformer states are less than 30, which are less than the data samples of the previous five types of states, resulting in a low accuracy rate of XGBoost machine training. The evaluation test results are shown in Table 2: the average accuracy of the evaluation of low-energy discharge, partial discharge, and low-temperature overheating is 77.92%, 84.30%, and 79.96%, respectively; while the average accuracy of the evaluation of the remaining five states is 94.04%, 88.92%, 91.95%, 91.18%, and 89.29%, respectively; thus, the remaining five types of states have a higher accuracy rate in the case of sufficient training samples. Under the premise of ensuring the number of training samples, XGBoost also has better performance in transformer state type evaluation.
[0110] Table 2 XGBoost state evaluation accuracy under default parameters
[0111]
[0112] The pre-processed DGA data is divided into non-coding ratios as feature parameters input into XGBoost, and the IPSO algorithm is used to automatically optimize multiple key parameters such as n_estimators, learning_rate, min_child_weight, and max_depth, and the particle fitness is the XGBoost state evaluation accuracy. In order to intuitively represent the performance of IPSO algorithm in optimizing XGBoost hyperparameters, the optimization effects of PSO and IPSO on XGBoost are compared, the initial population size popsize is 20, the optimization dimension dim is 5, and the iteration number Tmax is 100, and the optimization comparison results of traditional PSO and IPSO are shown in Figure 5 , and the process of model parameter learning rate optimization is shown in Figure 6 .
[0113] From Figure 5It can be seen that the IPSO algorithm falls into local optimum 3 times, and the accuracy reaches the maximum value of 95.57% at the 23th iteration. Compared with the traditional PSO algorithm, the IPSO algorithm is effectively improved in terms of getting rid of local optimum and optimization ability. From Figure 6 It can be seen that with the automatic optimization of the model parameter learning rate, the accuracy is gradually increasing in continuous training and iteration. Among them, when the learning rate is 0.431 and 0.605, the accuracy of the model evaluation reaches the peak. Combined with Figure 5 、 Figure 6 , it is necessary and effective to optimize the hyperparameters.
[0114] The four common models in machine learning: random forest RF, support vector machine SVM, k-neighbor algorithm KNN and logistic regression LR, the present application compares the machine training and evaluation results by using the same set of data sets with the above four machine learning models. It can be seen from Table 3 that the XGBoost proposed in the present application is superior to the other four models in terms of the accuracy of state evaluation and the stability of the evaluation results of different types of transformer states. In order to ensure the universality of the experimental results, the above classification models are trained and evaluated again by using the remaining four data sets, as shown in Table 4.
[0115] Table 3 Comparison of evaluation results of various models
[0116]
[0117] Table 4 Repeated training results of various models
[0118]
[0119] According to the comparison results, the intelligent evaluation model proposed in the present application has the highest accuracy and stability in evaluating the transformer state among the five machine learning models.
[0120] After the detection of the embodiments, the following conclusions are made:
[0121] (1) When using non-ratio encoding as feature input, compared with the traditional machine learning classification models RF, SVM, KNN and LR algorithm, XGBoost performs better and has more stable evaluation results in the evaluation of 8 typical transformer states.
[0122] (2) The global optimization ability of PSO algorithm is improved by dynamically adjusting the inertia weight and acceleration factor. The simulation results show that the optimization ability and the ability to get rid of local optimum of IPSO algorithm are better than those of the traditional PSO algorithm.
[0123] (3) The IPSO algorithm is applied to the hyperparameter optimization of XGBoost, so that the state evaluation accuracy of XGBoost is improved, and the model evaluation performance is more stable.
[0124] In conclusion, the intelligent evaluation method of the oil-immersed transformer state provided by the application improves the state evaluation performance of the XGBoost classification model, enhances the reliability of the evaluation result, effectively solves the problems of slow particle swarm optimization speed, optimization falling into local optimum, and the like, shortens the machine training and learning time, and improves the precision and stability of the transformer state evaluation.
[0125] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application.
Claims
1. A smart assessment method for the condition of an oil-immersed transformer, characterized in that, Includes the following steps: S1. Collect data from online transformer monitoring and oiling experiments, statistically analyze the composition and content of gases in transformer oil, and establish a DGA sample dataset. S2. Perform data filtering on the DGA sample data to remove invalid and interfering data; S3. The DGA data is preprocessed using the uncoded ratio method to characterize, normalize, and standardize the sample data; S4. Determine the feature dimension and encode the corresponding transformer state as the evaluation result of the model output; S5. Balance the data samples, group them, and randomly select 80% of the sample data as the training set, and the rest as the test set; S6. Establish a transformer evaluation model based on XGBoost; The process of establishing the XGBoost-based transformer evaluation model in step S6 is as follows: S61. Constructing the XGBoost transformer evaluation model F = {f(x) = ω q (x)}q: R m → T, ω ∈ R T (2) In equation (1), This represents the final prediction result when training up to the Kth Classification and Regression Tree (CART), where each f... k The function corresponds to a CART tree; x i Let represent the i-th sample; in equation (2), q and ω represent the structure vector and leaf weights of CART; R represents the space of the regression tree; m represents the number of features; T represents the number of leaf nodes; F is the set of all CARTs; x represents a certain sample; φ represents x i and Functional relationship; ω q This represents the predicted value of the CART classifier for classifying the sample; After the sample input, the structure vector q is directed to the corresponding leaf label according to the CART classification, each leaf label occupies different weights, ω q is the prediction value of the CART classifier for sample classification; Each CART classifier in the set independently classifies the input sample into a leaf label. Finally, the XGBoost model sums the predictions of all CART classifiers to obtain the prediction result. S62. Establish the objective function After inputting sample i, the regularization objective function for the k-th iteration is: In the formula, L represents the original loss function; l represents the cumulative statistics of the prediction results; y represents the prediction result for the i-th sample; i Ω represents the model prediction from the previous round; γ and λ are constants constraining the degree of regularization; T represents the number of leaf nodes and ω. 2 Indicates the leaf weight; S63, Optimization of the Objective Function The prediction at the kth iteration is equal to the prediction at the (k-1)th iteration plus f k The objective function is further written as: Applying the second-order Taylor expansion to the objective function: wherein L (k) represents the objective function using the second order Taylor expansion; g i and h i are the first and second order derivatives of the loss function, respectively; The constant part cannot be optimized; the optimizable part of the function is: In the formula, This represents the optimizable objective function after ignoring the constant part; Each sample will eventually be classified to a leaf node. Samples from the same leaf node are recombined, and the loss functions of each sample are summed. The process is as follows, let the leaf node be numbered j: In formula (8), ω j denotes the weight of the jth leaf node; I j is a set of numbers for all samples: I j ={i|q(x i )=j} (9) The objective function is rewritten as a quadratic function of the predicted values ω at the leaf nodes. Therefore, the optimal ω and objective function values can be directly solved using the vertex formula: In the formula, For ω j The optimal solution; Step S64: Introduce the split node algorithm In XGBoost training, a greedy algorithm is introduced to find the best split point more simply and quickly. When the amount of sample data increases, it puts a lot of demands on memory. The greedy algorithm will consume a lot of testing time to traverse all possible split points. At this time, an approximation algorithm is introduced to replace the greedy algorithm. S7. An IPSO-XGBoost evaluation model is proposed by using an improved particle swarm optimization to optimize the key hyperparameters of XGBoost. Step S7, which involves improving the key hyperparameters of XGBoost using particle swarm optimization, proposes the following specific process for the IPSO-XGBoost evaluation model: S71, Particle Swarm Optimization Algorithm Design In the Particle Swarm Optimization (PSO) algorithm, the optimization direction of a particle (potential solution) in the D-dimensional target search space (solution space) is influenced by three factors: the velocity of the previous iteration, individual cognition, and swarm orientation. Individual cognition refers to the particle adjusting its optimization direction for the next iteration based on its own learning; swarm orientation refers to the particle adjusting its optimization direction for the next iteration based on the optimal action path of the swarm. Therefore, parameters such as inertia weight and acceleration factor are introduced to describe the velocity and position updates of the i-th particle. V ij (k+1)=ωV ij (k)+c1r1(k)[p ij (k)-x ij (k)]+c2r2(k)[p gj (k)-x ij (k)]j∈D (12) x ij (k+1)=x ij (k)+V ij (k+1)j∈D (13) In the formula, V ij (k) represents the j-dimensional velocity component of the i-th particle in the k-th iteration; ω is the inertia weight of the particle iteration; c1 and c2 are the acceleration factors of the particle iteration; r1 and r2 represent random numbers between [0,1]; p ij (k) represents the j-dimensional component of the local optimal position of the i-th particle in the k-th iteration; x ij (k) represents the j-dimensional position component of the i-th particle in the k-th iteration; p gj (k) represents the j-dimensional component of the overall optimal position of the i-th particle in the k-th iteration; S72, Improved Particle Swarm Optimization Algorithm For the three parameters in equations (12) and (13), ω describes the "inertia" of the particle. In the early stage of iteration, ω should be larger to ensure that the particle can fly independently and search the space sufficiently. In the later stage of iteration, ω should be smaller to learn more from other particles. c1 and c2 describe the effects of the local optimal position and the global optimal position, respectively. In the early stage, c1 should be larger, and in the later stage, c2 should be larger. This is to balance the global search capability and local search capability of the particle. In view of this, the traditional particle swarm algorithm is improved, and the improved particle swarm (IPSO) algorithm is proposed. The above parameters are improved by the following formula: ω(k)=ω max -(oh max -oh min (k / T) max ) 2 (14) c1(k)=c 10 -(c 10 -c 11 )(k / T max ) (15) c2(k)=c 20 -(c 20 -c 22 )(f / T max ) (16) In the formula, ω(k), c1(k), and c2(k) represent the inertial weight and acceleration factor of the particle in the kth iteration, respectively, and are used to replace ω, c1, and c2 in equation (12); ω max ω min These represent the maximum and minimum values of the inertial weight, respectively; c 10 c 11 and c 20 c 22 Let T represent the initial and final values of the acceleration factors c1 and c2, respectively; max Indicates the maximum number of iterations; S73. Establish the IPSO-XGBoost evaluation model The IPSO algorithm is introduced into the XGBoost transformer evaluation model established in step S6 to balance the ability of global search and local search during particle optimization. The IPSO algorithm is used to optimize multiple key hyperparameters of XGBoost simultaneously, and the IPSO-XGBoost evaluation model is established. S8. Import the training set into the IPSO-XGBoost evaluation model for machine training, and import the test set into the IPSO-XGBoost evaluation model to perform intelligent evaluation of the transformer state.
2. The intelligent assessment method for the condition of an oil-immersed transformer according to claim 1, characterized in that, In step S1, the composition of the gas in the transformer oil includes H2, CH4, C2H2, C2H4, C2H6, as well as various alkanes and unsaturated hydrocarbons.
3. The intelligent assessment method for the condition of an oil-immersed transformer according to claim 1, characterized in that, In step S2, the DGA sample dataset is filtered to remove invalid and interfering data, including but not limited to data with fewer than three components in the transformer oil and data whose gas content differs from the historical average by more than three orders of magnitude.
4. The intelligent assessment method for the condition of an oil-immersed transformer according to claim 1, characterized in that, Step S3, which uses a non-coded ratio method to preprocess DGA data, employs the following nine gas content ratios to form characteristic quantities: CH4 / H2, C2H4 / C2H2, C2H4 / C2H6, (CH4+C2H4) / (C1+C2), H2 / (H2+C1+C2), C2H4 / (C1+C2), CH4 / (C1+C2), C2H6 / (C1+C2), C2H2 / (C1+C2); Wherein, C1 is the sum of the contents of alkane gases represented by CH4; C2 is the sum of the contents of unsaturated hydrocarbon gases represented by C2H6, C2H4, and C2H2.
5. The intelligent assessment method for the condition of an oil-immersed transformer according to claim 1, characterized in that, The specific form of encoding the transformer state in step S4 is as follows: According to the DL / T 722-2014 guideline, the transformer status output corresponds to 8 status codes: Normal state is 0; arc discharge is 1; low energy discharge is 2; partial discharge is 3; high temperature overheating is 4; medium temperature overheating is 5; low temperature overheating is 6; high energy discharge is 7.
6. The intelligent assessment method for the condition of an oil-immersed transformer according to claim 1, characterized in that, In step S5, the method for balancing the data samples by grouping is as follows: the DGA data of different transformer states are classified and statistically analyzed, and then evenly distributed to the 5 groups according to the proportion to ensure the balance of data between the groups.
7. The intelligent assessment method for the condition of an oil-immersed transformer according to claim 1, characterized in that, The key hyperparameters in step S73 include, but are not limited to, the number of classifiers, the minimum sum of sample weights in the child nodes, the learning rate, the depth of the tree, and the seed for the random generation function.
8. The intelligent assessment method for the condition of an oil-immersed transformer according to claim 1, characterized in that, The specific process for step S8, which involves importing the training set into the IPSO-XGBoost evaluation model for machine training and importing the test set into the IPSO-XGBoost evaluation model to perform intelligent evaluation of the transformer state, is as follows: S81. Import the training set into the IPSO-XGBoost evaluation model and perform machine training; S82 and IPSO calculate the corresponding particle fitness for model hyperparameters; S83 and IPSO search the population space, update the best position of particles and the globally optimal particle, and further update the particle search speed, position and inertia weight; S84. Determine whether the model's evaluation result matches the actual state of the transformer. If not, repeat steps S82 and S83. If yes, output the optimal parameter particle to achieve automatic optimization of the model's hyperparameters. S85. Input the test set into the evaluation model after the hyperparameters have been optimized, perform transformer state evaluation, and output the model evaluation results. S86. Analyze the performance of the evaluation model based on the evaluation results.