Research On Power Transformer Fault Diagnosis And Prediction Based On Deep Learning

Power transformers play an important role in power systems,therefore,their operating states are closely related to the reliability of the power grid.With the construction of ubiquitous power internet of things,a sharing data platform has been established including power transformer state parameters,system load data and meteorological data.Under the condition of massive data accumulation,deep learning provides a more accurate and reliable data support on fault diagnosis and prediction than traditional models by its fine modeling in multi-dimension attributes and long-term time sequences.In terms of imbalanced class distribution among fault categories of power transformers which leads to low diagnostic accuracy rate and the large difference in recognition effectiveness between fault types of traditional machine learning methods,we propose a multi-level hierarchical power transformer fault diagnosis model according to hierarchical classification and ensemble learning.In essence,classifiers are hierarchically constructed for level-by-level diagnosis with consideration of the extent of imbalance on each level.The neural network classifier at first level extracts three generalized feature tags of normal,discharge fault and overheat fault for feature fusion with the original data input,to guide classification among 9 exact operation states.The classifier at second level adopts Easy Ensemble algorithm which generates balanced training subsets by under-sampling the majority class and synthesizes the sub-classifiers’ parameters for the ultimate classifier by parallel training.The prediction results given by the traditional methods on state parameters are not precise enough due to the lack of association mining or the long prediction time scale.In this paper,a dual-stage attention mechanism is introduced to long short-term memory network(LSTM)as the predictive model,to extract association rules between input features and time dependencies between history time points.The feature attention mechanism is used for mining the relationship among the target parameter and other states’ information,environment and operation data automatically,in order to correct results appropriately.The temporal attention mechanism selects key time points in history for information enhancement over basic LSTM model,stabilizing the performance over different prediction period.Based on the dissolved gas analysis theory in condition evaluation,a hybrid deep learning network is built by state parameter prediction module and operating state diagnostic module.The historical data of concentrations of dissolved gas,top oil temperature,and other related attributes are modeled to obtain precise predictions on dissolved gas concentration through the parameter prediction module.Then the operation state classifier calculates the probability distribution among each fault type by the predicted values.The operating state with the maximal confidence probability is selected as the future operation state of the transformer.The case study shows that equipment either in normal operation or in defective operation could be correctly claimed by this hybrid model.A software on power transformer state recognition is also developed.By combining and analyzing the multi-source information of power transformer through deep learning algorithms,practical functions such as data cleaning,condition evaluation,fault diagnosis and state prediction are applied.Historical data supplied by technology project is used for time series modeling and parameter,followed by state prediction with the latent defect probability distribution.This software provides a visual interface and reliable data support for state monitoring and smart condition evaluation.