Research On Transformer State Prediction And Fault Diagnosis Based On Neural Network And Attention Mechanism

As an important device in the power system,the safe operation of power transformers is related to the stability of the entire power grid.At present,the accident rate of on-site transformers being forced to shut down for maintenance due to abnormal conditions or faults is increasing.With the continuous increase of the overall power consumption of the society,the increase of transformer investment time,and the access of a large number of intelligent equipment and sensors,it has been difficult to process current multi-dimensional data information based on theoretical analysis and other causal relationship models in the past.With the rapid development of artificial intelligence,by searching for the inherent laws of the data itself,using artificial intelligence algorithms for data mining,it provides new solutions and technical means for the parameter prediction and fault diagnosis of power transformers.Aiming at the problem of weak generalization and low stability of the current trend prediction of dissolved gases in the oil of power transformers,this paper proposes a combined improved cyclic network model.This model is based on the latest cyclic network structure GRU model,improves the double-loop structure and incorporates the attention mechanism.The gate control structure of the GRU network can make it remember the trend of changes when processing long-term sequence data,while the double-loop structure and time attention mechanism can help the model be more sensitive to changes in data during training.Through actual case analysis,the prediction accuracy of the combined model is higher than that of the traditional gas prediction network,and it can more accurately track the trend of oil chromatogram changes when the gas content changes suddenly.In order to improve the stability and accuracy of transformer fault classification,a transformer fault classification model based on convolutional network is proposed.The use of convolution kernels to mine the connections between various gas parameters makes the model’s classification of different transformer faults clearer.Adding a pooling layer can make the model have lower stability and false alarm rate when processing field data sets,and avoid direct models from the past.Potential risks caused by deleting the default data.Validation of actual transformer fault data sets shows that,compared with traditional neural network or multilayer perceptron models,the fault diagnosis model in this paper has a relatively higher accuracy rate and is more stable in the face of different data sets.Aiming at the black box problem of the neural network model itself and the need to find out the cause of the transformer fault on the actual site,the self-attention mechanism is introduced into the classification model,and the model’s attention weight is visualized to explain the reason for the model’s judgment.The analysis of the existing results shows that the attention mechanism has a certain meaning in the interpretation model for transformer fault judgment.The research content of this paper has enriched the current parameter prediction and fault diagnosis methods of power transformers,and has made certain progress in the versatility and stability of model prediction and judgment.At the same time,it has certain value for the interpretation and visualization of transformer fault judgment.