Prediction of Transformer Oil Temperature Based on Feature Selection and Deep Neural Network

The prediction of transformer oil temperature is a time series prediction task, which is of great significance for the maintenance of substation equipment. In the process of multivariate prediction, not all variables are strongly correlated with the prediction of the transformer oil temperature. This paper proposes a transformer oil temperature prediction method based on Feature Selection and Smooth Residual blocks, named FSSR. A two-stage feature selection method is applied to select the related features. Residual connection is applied to convolutional neural network in smooth residual blocks. Experimental study is performed on the ETTh1 and ETTh2 datasets. With different number of selected features, the best one can reduce the average MSE (mean squared error) by at least 20% than the worst one. Compared to the baseline models, FSSR has the best accuracy. Compared with LSTM (Long Short-Term Memory network), the average MSE of FSSR decreased by 27.62% on the ETTh1 dataset, and 39.34% on the ETTh2 dataset.