Research On Transformer Operation State Prediction Based On BiLSTM And KELM Combined Model

As an important part of the power system,the inspection of the condition of the transformer is essential to maintain the stable operation of the power system.The increasing number of abnormal operating conditions and faults in transformers has led to an increase in the number of cases where work has to be stopped for maintenance in the field.In the past,theoretical analysis methods based on cause-and-effect models have been difficult to adapt to the current needs of processing diverse data and information on transformers.With the continuous advancement of artificial intelligence technology,the use of data mining methods combined with advanced artificial intelligence algorithms to find the inherent laws of power transformer data has brought new tools and solutions for condition prediction and fault diagnosis of power transformers.In terms of prediction,the most commonly used neural network model for predicting gas concentration sequences is the long and short-term memory network model.Although long short-term memory networks are capable of capturing correlations between sequences,they may suffer from a representation bottleneck and are unable to adequately extract and represent the effective features in sequence data.To improve the performance of the LSTM model,a new prediction model,CNNBiLSTM-Attention,was constructed,i.e.adding a reverse LSTM layer to the LSTM model,while using a convolutional neural network to improve the utilisation of the feature data and adding an attention mechanism to better identify important features.Using the prediction model proposed in this paper,experiments are conducted on seven typical feature gases and the experimental results show that the model is efficient and accurate in predicting the future content of the feature gases.In fault diagnosis,this paper proposes a sparrow search algorithm combined with a kernel limit learning machine to diagnose the operating condition of transformers using the "uncoded ratio" of seven characteristic gases as the characteristic quantity.The method uses principal component analysis to reduce the dimensionality of the data in order to retain the non-linear information of the original features.The sparrow search algorithm is used to find the optimal regularisation coefficients and kernel parameters,which effectively improves the global search capability and convergence speed of the model,and also solves the problem that the model tends to fall into local optimal solutions,further improving the diagnostic effect of the kernel limit learning machine model.Through example verification,the proposed method has high diagnostic accuracy and can determine various faults occurring in transformers by the given characteristic gas concentrations.Finally,this paper designs an intelligent system that can be used for online monitoring and condition prediction of transformers,and validates the system solution through effective testing of practical cases,fully proving the reliability and practicality of the system,which can effectively meet the actual needs of engineering.