Transformer dissolved gas concentration prediction method based on D-S evidence theory optimization ensemble learning

Abstract

The invention relates to the technical field of transformer dissolved gas concentration prediction, in particular to a transformer dissolved gas concentration prediction method based on D-S evidence theory optimization ensemble learning. Comprising the following steps: generating sub-data sets, training a base learner for each sub-data set to obtain a plurality of independent models, predicting a to-be-predicted sample by using the trained base learner to obtain a prediction result, calculating a basic probability allocation for the prediction result of each base learner, and calculating a prediction result of each base learner; fusing the prediction results of all the base learners by using a D-S evidence theory and generating a final prediction value, and predicting the concentration of the dissolved gas of the transformer based on an integrated learning model optimized by the D-S evidence theory. The method is based on a Bagging method in ensemble learning, a D-S evidence theory is used for fusion of a base learner, and the transformer dissolved gas concentration prediction accuracy of an ensemble learning model is improved; and a sequence least square programming algorithm is adopted, so that complex constraint conditions can be conveniently processed, and the convergence speed and the calculation efficiency are improved.

Description

Technical Field

[0001] The invention relates to the technical field of transformer dissolved gas concentration prediction, and in particular to a transformer dissolved gas concentration prediction method based on DS evidence theory optimized ensemble learning. Background Art

[0002] Transformers play a vital role in the power grid system, mainly responsible for voltage conversion and power distribution. It converts the high-voltage power generated by the power plant into a voltage level suitable for transmission and distribution, thereby ensuring the efficient transmission of power. At the same time, the transformer provides electrical isolation to protect the safety of equipment and personnel. However, transformers may face various faults during operation, so condition monitoring is particularly important, and in transformer condition monitoring, the prediction of dissolved gas concentration is even more critical. Changes in the concentration of dissolved gas in transformer oil can reflect its internal fault conditions, such as overheating or breakdown. By monitoring and predicting the gas concentration, potential problems can be warned in advance, and reasonable maintenance and overhaul plans can be formulated to reduce unnecessary downtime.

[0003] Therefore, there have emerged a transformer fault diagnosis method and system based on oil dissolved gas analysis disclosed in the invention patent with publication number CN118779810A and a power transformer fault early warning method and system disclosed in the invention patent with publication number CN117829822B, which predict the dissolved gas concentration of the transformer.

[0004] However, during use, it was found that since the prediction of dissolved gas concentration in the transformer is a nonlinear problem, the traditional statistical method is generally ineffective in dealing with this problem, and the deep learning method has too high data requirements and complex calculations. The data in the power grid is difficult to meet its needs, and it is easy to bring excessive computing pressure to the power grid system, resulting in poor practicality. Therefore, there is an urgent need for a transformer dissolved gas concentration prediction method based on DS evidence theory optimized integrated learning to improve the above problems. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a Bagging method based on ensemble learning, and uses DS evidence theory for the fusion of base learners, which improves the prediction accuracy of transformer dissolved gas concentration of the ensemble learning model, and can better handle the uncertainty brought by the sub-data set, and enhances the robustness of the model to noise data; secondly, a sequential least squares programming algorithm is adopted to gradually update the decision variables in an iterative manner, and finally find the optimal solution that meets all constraints, which is convenient for processing complex constraints and improves the convergence speed and computational efficiency of a transformer dissolved gas concentration prediction method based on DS evidence theory optimized ensemble learning.

[0006] A transformer dissolved gas concentration prediction method based on DS evidence theory optimization integrated learning of the present invention comprises the following steps:

[0007] S1. Generate sub-datasets, that is, randomly extract multiple independent sub-datasets from the original data set with replacement. In the Bagging method, random data samples in the training set are selected by replacement method;

[0008] S2. Train a base learner for each sub-data set to obtain multiple independent models, that is, train a decision tree model for each sub-data set to obtain multiple independent decision tree models;

[0009] S3, using the trained base learner to predict the sample to be predicted to obtain the prediction result, that is, using the multiple decision tree models obtained in S2 to predict the sample to be predicted at the same time to obtain the prediction results of different decision tree models respectively;

[0010] S4, calculating the basic probability distribution for the prediction results of each base learner, the initial basic probability distribution value is the average value, and then the basic probability distribution value is continuously updated according to the optimization algorithm with the goal of minimizing the mean square error of the prediction results;

[0011] S5. Use DS evidence theory to fuse the prediction results of all base learners and generate the final prediction value. After obtaining the fused prediction result, evaluate the result. When the mean square error of the prediction result tends to converge, enter S6. Otherwise, return to S4 to recalculate the basic probability distribution.

[0012] S6. Based on the ensemble learning model optimized by DS evidence theory, the concentration of dissolved gas in transformer is predicted, and the prediction results are obtained and compared with the actual results to analyze the accuracy of the prediction of this method;

[0013] S7. After comparison, if the prediction accuracy meets the required standard, it will be directly applied to the prediction of dissolved gas concentration in actual transformers.

[0014] Preferably, the training process of the decision tree model in S2 includes:

[0015] Select split features, traverse all features and their possible split points, and select the optimal feature for splitting based on information gain;

[0016] The tree is recursively constructed, that is, the feature selection and node splitting process is repeated for each subset until the stopping condition is met, where the stopping condition includes that the depth of the tree exceeds the preset maximum value, the number of samples in the node is lower than the set threshold, and further splitting cannot significantly improve the information gain.

[0017] Preferably, when predicting in S3, starting from the root node, traverse downward along the decision tree according to the feature value, each node will be judged according to the selected feature and threshold, and if the feature value meets a certain condition, it will move downward along one path, otherwise it will move along another path until it reaches a leaf node and returns the predicted value of the leaf node.

[0018] Preferably, the specific method of calculating the basic probability distribution for the prediction result of each base learner in S4 includes:

[0019] During the optimization process, the sequential least squares programming algorithm is used as the optimization algorithm, and its objective function satisfies the relationship: where Y i In this application, it is the actual result value, y i is the aforementioned fusion prediction value.

[0020] Preferably, in the optimization process, the rule for updating the basic probability distribution value in each iteration satisfies the relationship: k+1 =m k +α k *d k , where α k is the step length, d k To calculate the direction, the core is to construct the objective function of the mean square error, set the constraint of the sum of the basic probability distribution values, and iteratively update the basic probability distribution values ​​to find the optimal solution.

[0021] Preferably, in S5, the specific method of using DS evidence theory to fuse the prediction results of all base learners and generate the final prediction value includes:

[0022] The basic probability distribution in DS evidence theory is a representation of the credibility of a hypothesis or event. For all possible events, the sum of the basic probability distribution is 1. After applying DS evidence theory to the fusion of the prediction results of the Bagging method, the calculated fusion prediction value satisfies the relationship: Among them, m i Assign a value to the base probability of the i-th base learner, f i is the i-th base learner, f i (x) is the prediction result of the i-th base learner. The application of DS evidence theory in the code provides an effective result fusion mechanism for the Bagging model. By weighting the predictions of different models, the information can be better integrated and the uncertainty of the prediction can be reduced. By flexibly setting the trust allocation value, the prediction performance of the model can be improved. In particular, it effectively solves the problem of unstable fusion prediction results caused by the randomness of sub-datasets in ensemble learning.

[0023] Preferably, the specific method for predicting the dissolved gas concentration of the transformer based on the integrated learning model optimized by DS evidence theory in S6 includes:

[0024] The data set to be predicted is divided into sub-datasets, and each sub-dataset is predicted by the base learner. After the prediction results of each base learner are obtained, the prediction results of all base learners are fused using the DS evidence theory based on the optimal basic probability distribution obtained through multiple iterations of S4 and S5 to obtain the prediction results. Finally, the fused prediction results are compared with the actual values.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. Based on the Bagging method in ensemble learning and using DS evidence theory for the fusion of base learners, the prediction accuracy of transformer dissolved gas concentration of the ensemble learning model is improved, and the uncertainty brought by sub-data sets can be better handled, thus enhancing the robustness of the model to noise data;

[0027] 2. Adopt the sequential least squares programming algorithm to gradually update the decision variables in an iterative manner, and finally find the optimal solution that meets all constraints, which is convenient for handling complex constraints and improves the convergence speed and computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of a transformer dissolved gas concentration prediction method based on DS evidence theory optimization integrated learning of the present invention;

[0029] Figure 2 It is a radar chart comparing the accuracy of the experimental prediction results of the present invention. DETAILED DESCRIPTION

[0030] In order to facilitate understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0031] Example

[0032] like Figure 1 As shown, a transformer dissolved gas concentration prediction method based on DS evidence theory optimization ensemble learning includes the following steps:

[0033] S1. Generate sub-datasets, that is, randomly extract multiple independent sub-datasets from the original data set with replacement. In the Bagging method, random data samples in the training set are selected by replacement method;

[0034] S2. Train a base learner for each sub-data set to obtain multiple independent models, that is, train a decision tree model for each sub-data set to obtain multiple independent decision tree models, wherein the training process of the decision tree model includes:

[0035] Select split features, traverse all features and their possible split points, and select the optimal feature for splitting based on information gain;

[0036] Recursively build the tree, that is, repeat the feature selection and node splitting process for each subset until the stopping condition is met, where the stopping condition includes that the depth of the tree exceeds the preset maximum value, the number of samples in the node is lower than the set threshold, and further splitting cannot significantly improve the information gain;

[0037] S3, use the trained base learner to predict the sample to be predicted to obtain the prediction result, that is, use the multiple decision tree models obtained in S2 to predict the sample to be predicted at the same time to obtain the prediction results of different decision tree models. When predicting, start from the root node and traverse down along the decision tree according to the feature value. Each node will be judged according to the selected feature and threshold. If the feature value meets a certain condition, it will move down along one path, otherwise it will move along another path until it reaches a leaf node and returns the prediction value of the leaf node;

[0038] S4. Calculate the basic probability distribution for the prediction results of each base learner. The initial basic probability distribution value is the average value. Then, the basic probability distribution value is continuously updated according to the optimization algorithm with the goal of minimizing the mean square error of the prediction results. The specific method includes:

[0039] During the optimization process, the sequential least squares programming algorithm is used as the optimization algorithm, and its objective function satisfies the relationship: where Y i In this application, it is the actual result value, y i is the aforementioned fusion prediction value; the rule for updating the basic probability distribution value in each iteration satisfies the relationship: k+1 =m k +α k *d k , where α k is the step length, d k The core of the calculation is to construct the objective function of mean square error, set the constraint of the sum of basic probability distribution values, and iterate and update the basic probability distribution values ​​to find the optimal solution;

[0040] S5. Use DS evidence theory to fuse the prediction results of all base learners and generate the final prediction value. After obtaining the fused prediction result, evaluate the result. When the mean square error of the prediction result tends to converge, enter S6. Otherwise, return to S4 to recalculate the basic probability distribution. The specific method includes:

[0041] The basic probability distribution in DS evidence theory is a representation of the credibility of a hypothesis or event. For all possible events, the sum of the basic probability distribution is 1. After applying DS evidence theory to the fusion of the prediction results of the Bagging method, the calculated fusion prediction value satisfies the relationship: Among them, m i Assign a value to the base probability of the i-th base learner, f i is the i-th base learner, f i (x) is the prediction result of the i-th base learner. The application of DS evidence theory in the code provides an effective result fusion mechanism for the Bagging model. By weighting the predictions of different models, the information can be better integrated and the uncertainty of the prediction can be reduced. By flexibly setting the trust allocation value, the prediction performance of the model can be improved. In particular, it effectively solves the problem of unstable fusion prediction results caused by the randomness of sub-datasets in ensemble learning.

[0042] S6. Based on the ensemble learning model optimized by DS evidence theory, the concentration of dissolved gas in transformer is predicted, and the prediction results are obtained and compared with the actual results to analyze the accuracy of the prediction method. The specific methods include:

[0043] The data set to be predicted is divided into sub-datasets, and each sub-dataset is predicted by the base learner. After the prediction results of each base learner are obtained, the prediction results of all base learners are fused by DS evidence theory based on the optimal basic probability distribution obtained by multiple iterations of S4 and S5. Finally, the fused prediction results are compared with the actual values;

[0044] S7. After comparison, if the prediction accuracy meets the required standard, it will be directly applied to the prediction of dissolved gas concentration in actual transformers.

[0045] Experimental example

[0046] The experiment was conducted based on 21 substation continuous time gas concentration change data, including four gases: H2, C2H6, C2H4, and CH4. The first 20 data were used as training sets to train an ensemble learning model based on DS evidence theory optimization, and then the trained model was used to predict the concentration of each gas at the next moment.

[0047] Based on the above experimental settings, the experiment was conducted and the results were shown in Table 1. This application defines that the prediction accuracy of each model for a certain gas satisfies the relationship:

[0048]

[0049] And the overall prediction accuracy of each model satisfies the relationship:

[0050]

[0051] Where C is the number of gas categories, is the actual concentration of the i-th gas, Y i Predicted values ​​for the model;

[0052] Table 1 Prediction results

[0053]

[0054] Therefore, we calculated the prediction accuracy of ensemble learning and ensemble learning based on DS evidence theory optimization according to the above formula. In H2 gas prediction, the prediction accuracy of ensemble learning is about 98.7%, and the prediction accuracy of ensemble learning based on DS evidence theory optimization is about 99%; in C2H6 gas prediction, the prediction accuracy of ensemble learning is about 98.1%, and the prediction accuracy of ensemble learning based on DS evidence theory optimization is about 99.7%; in C2H4 gas prediction, the prediction accuracy of ensemble learning is about 99.6%, and the prediction accuracy of ensemble learning based on DS evidence theory optimization is about 99.9%; in CH4 gas prediction, the prediction accuracy of ensemble learning is about 97.7%, and the prediction accuracy of ensemble learning based on DS evidence theory optimization is about 96.8%. As shown in Table 2, although the prediction accuracy of ensemble learning based on DS evidence theory optimization is slightly lower than that of ensemble learning in the prediction of CH4 gas, the overall accuracy has been improved;

[0055] Table 2 Prediction accuracy results

[0056]

[0057] like Figure 2 As shown, for the convenience of intuitive comparison, the present application draws a radar chart based on the data in Table 2. From the size of the two values ​​on the single axis and the proportion of the two areas, it can be more intuitively seen that the prediction accuracy of the DS evidence theory optimized ensemble learning is lower than that of the ensemble learning only in the prediction of CH4 gas, and the other three gases and the overall prediction accuracy are significantly higher than the ensemble learning method. Therefore, it can be proved that the DS evidence theory optimized ensemble learning transformer dissolved gas concentration prediction method proposed in this application is effective, as well as the improvement of the accuracy and robustness of the original ensemble learning method.

[0058] Since the objective function of the sequential least squares programming algorithm included in the DS evidence theory optimization integrated learning transformer dissolved gas concentration prediction method proposed in this application is to minimize the overall mean square error, this application finally calculates the mean square error value of the prediction results of the DS evidence theory optimization integrated learning and integrated learning methods based on the experimental results, and verifies the effectiveness of the optimization algorithm from the side. Among them, the calculation method of the mean square error satisfies the relationship: The results calculated according to the above formula are shown in Table 3. It can be seen that the overall mean square error of the prediction results based on the DS evidence theory optimization integrated learning is only about 22% of the mean square error of the integrated learning prediction results, indicating that the DS evidence theory-based optimization integrated learning method has greatly reduced the mean square error of the overall prediction results. At the same time, it can also be concluded that the stability of the DS evidence theory-optimized integrated learning transformer dissolved gas concentration prediction method proposed in this application and the effectiveness of the sequential least squares planning optimization algorithm;

[0059] Table 3 Prediction mean square error results

[0060]

[0061] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A transformer dissolved gas concentration prediction method based on DS evidence theory optimization ensemble learning, characterized in that: The following steps are involved: S1. Generate sub-datasets, that is, randomly extract multiple independent sub-datasets from the original data set with replacement. In the Bagging method, random data samples in the training set are selected by replacement method; S2. Train a base learner for each sub-data set to obtain multiple independent models, that is, train a decision tree model for each sub-data set to obtain multiple independent decision tree models; S3, using the trained base learner to predict the sample to be predicted to obtain the prediction result, that is, using the multiple decision tree models obtained in S2 to predict the sample to be predicted at the same time to obtain the prediction results of different decision tree models respectively; S4, calculating the basic probability distribution for the prediction results of each base learner, the initial basic probability distribution value is the average value, and then the basic probability distribution value is continuously updated according to the optimization algorithm with the goal of minimizing the mean square error of the prediction results; S5. Use DS evidence theory to fuse the prediction results of all base learners and generate the final prediction value. After obtaining the fused prediction result, evaluate the result. When the mean square error of the prediction result tends to converge, enter S6. Otherwise, return to S4 to recalculate the basic probability distribution. S6. Based on the ensemble learning model optimized by DS evidence theory, the concentration of dissolved gas in transformer is predicted, and the prediction results are obtained and compared with the actual results to analyze the accuracy of the prediction of this method; S7. After comparison, if the prediction accuracy meets the required standard, it will be directly applied to the prediction of dissolved gas concentration in actual transformers.

2. A transformer dissolved gas concentration prediction method based on DS evidence theory optimization ensemble learning as claimed in claim 1, characterized in that: The training process of the decision tree model in S2 includes: Select split features, traverse all features and their possible split points, and select the optimal feature for splitting based on information gain; The tree is recursively constructed, that is, the feature selection and node splitting process is repeated for each subset until the stopping condition is met, where the stopping condition includes that the depth of the tree exceeds the preset maximum value, the number of samples in the node is lower than the set threshold, and further splitting cannot significantly improve the information gain.

3. The transformer dissolved gas concentration prediction method based on DS evidence theory optimization ensemble learning as claimed in claim 1 is characterized in that: When predicting in S3, starting from the root node, traverse downward along the decision tree according to the feature value, each node will be judged according to the selected feature and threshold, and if the feature value meets a certain condition, it will move downward along one path, otherwise it will move along another path until it reaches a leaf node and returns the predicted value of the leaf node.

4. A transformer dissolved gas concentration prediction method based on DS evidence theory optimization ensemble learning as claimed in claim 1, characterized in that: The specific method of calculating the basic probability distribution for the prediction results of each base learner in S4 includes: During the optimization process, the sequential least squares programming algorithm is used as the optimization algorithm, and its objective function satisfies the relationship: where Y i In this application, it is the actual result value, y i is the aforementioned fusion prediction value.

5. A transformer dissolved gas concentration prediction method based on DS evidence theory optimization ensemble learning as claimed in claim 4, characterized in that: During the optimization process, the rule for updating the basic probability distribution value in each iteration satisfies the relationship: k+1 =m k +α k *d k , where α k is the step length, d k To calculate the direction, the core is to construct the objective function of the mean square error, set the constraint of the sum of the basic probability distribution values, and iteratively update the basic probability distribution values ​​to find the optimal solution.

6. The transformer dissolved gas concentration prediction method based on DS evidence theory optimization ensemble learning as claimed in claim 1 is characterized in that: In S5, the specific method of using DS evidence theory to fuse the prediction results of all base learners and generate the final prediction value includes: The basic probability distribution in DS evidence theory is a representation of the credibility of a hypothesis or event. For all possible events, the sum of the basic probability distribution is 1. After applying DS evidence theory to the fusion of the prediction results of the Bagging method, the calculated fusion prediction value satisfies the relationship: Among them, m i Assign a value to the base probability of the i-th base learner, f i is the i-th base learner, f i (x) is the prediction result of the i-th base learner. The application of DS evidence theory in the code provides an effective result fusion mechanism for the Bagging model. By weighting the predictions of different models, the information can be better integrated and the uncertainty of the prediction can be reduced. By flexibly setting the trust allocation value, the prediction performance of the model can be improved. In particular, it effectively solves the problem of unstable fusion prediction results caused by the randomness of sub-datasets in ensemble learning.

7. The transformer dissolved gas concentration prediction method based on DS evidence theory optimization ensemble learning as claimed in claim 1 is characterized in that: The specific method for predicting the dissolved gas concentration of the transformer based on the integrated learning model optimized by DS evidence theory in S6 includes: The data set to be predicted is divided into sub-datasets, and each sub-dataset is predicted by the base learner. After the prediction results of each base learner are obtained, the prediction results of all base learners are fused using the DS evidence theory based on the optimal basic probability distribution obtained through multiple iterations of S4 and S5 to obtain the prediction results. Finally, the fused prediction results are compared with the actual values.

Images