基于注意力—门控循环—时间卷积网络的变压器油温预测OA

Top oil temperature is an important indicator for evaluating thermal characteristics of power transformer.Accurately predicting the variation trend of the top oil temperature plays a crucial role in assessing the load capaci-ty of the transformer and promptly identifying potential thermal faults.In view of such issues as low accuracy,cum-bersome feature selection process,and incorrect use of information in top oil temperature prediction,a transformer top oil temperature prediction method(Attention-GRU-TCN)is proposed based on temporal convolutional network(TCN),attention mechanism,and gated recurrent unit(GRU).First,TCN is introduced into the convolution to en-sure causality in the prediction process and reduce the interference of future data on the prediction process.Then,the introduction of the attention mechanism assigns higher weights to key data features,replacing the complex and time-consuming data filtering process,thereby improving prediction accuracy.Finally,the proposed method is ex-perimentally validated at a company against traditional top oil temperature prediction models(e.g.,RNN,LSTM,GRU,BiLSTM,andTransformer,ect).The results show that compared with the RNN model,LSTM model,GRU model,BiLSTM model and Transformer model,the Attention-GRU-TCN model achieves a decrease in mean abso-lute error(eMAE)by 23.69%,15.88%,7.77%,29.35%,and 14.51%respectively,a decrease in mean squared error(eMSE)by 45.79%,31.34%,12.31%,43.32%,and 38.85%respectively;and an increase in the coefficient of determi-nation(R2)by 6.0%,3.10%,0.79%,12.30%,and 4.57%respectively.It is shown that the method has higher predic-tion accuracy for transformer top oil temperature,which is conducive to providing a reference for early warning of po-tential thermal faults and analyzing the load capacity of transformers.