Research On Early Warning Of Transformer Top Oil Temperature Anomalies Based On Machine Learning

Oil-immersed transformers are the most widely used transformers,and their top oil temperature state research have important production and research value.The current top oil temperature anomaly related technology has the characteristics of high equipment cost and complicated technical operation.According to the current situation of insufficient data of abnormal oil temperature in the State Grid transformer,this paper proposes a new two-stage oil temperature anomaly warning strategy.Among them: the first stage is the top oil temperature anomaly labeling based on semi-supervised learning,and the second stage is the abnormal oil temperature warning algorithm based on LSTM recurrent neural network.The related research is summarized as follows.The first stage: firstly,using k-means algorithm to find the different working conditions clustering of oil temperature unlabeled training set;Then,based on the oil temperature anomaly labeling threshold,the oil temperature anomaly labeling is performed on the data of each oil temperature interval under each cluster cluster of the oil temperature unlabeled training set;Secondly,based on the rule extraction and practical application requirements,the oil temperature training after the abnormal annotation is concentrated into the random forest model.After the training,the oil temperature anomaly decision rule and the oil temperature anomaly classifier can be obtained.Finally,the model is carried out in the manual marking test set.The test,if the specified accuracy rate and recall rate requirements are met,the goal is reached,otherwise the threshold is optimized by the threshold optimization algorithm and relabeled and tested.The second stage: firstly,the oil temperature training set labeled with the abnormal label in the first stage is obtained through the data set conversion algorithm to obtain the training data set,and is imported into the LSTM recurrent neural network model;Complete the training of the model relying on the feature extraction of the hidden layer and the LSTM recurrent neural network have the ability to learn long-term and short-term dependencies;finally,the model completed by the training can be used to predict the oil temperature anomaly.Since the LSTM recurrent neural network has the ability to learn long-term and short-term dependencies,the prediction effect of using LSTM recurrent neural network in the oil temperature anomaly prediction scenario is better than the traditional data mining method.In this paper,the proposed algorithm is verified by multi-angle comparison experiments,which proves its effectiveness and feasibility.At the same time,the scheme of this paper is put into production measurement in the oil-immersed transformer of S city,which further proves the feasibility and effectiveness of the above scheme.This article is based on semi-supervised learning,using both labeled and unlabeled samples,reducing the cost of sample annotation and achieving good results.At the same time,no additional testing equipment is needed,it can be applied to different types of transformers,and has good generalization.