Abstract: The service life of a transformer is directly related to its insulation performance. For UHV converter transformers, oil temperature forecasting can be used as an important basis for evaluating its insulation performance. In this paper, a method for top-oil temperature forecasting which combined long-short term memory networks (LSTM), self-attention mechanism (SA) and gated recurrent unit (GRU) was proposed, aiming to enhance temperature forecasting accuracy of converter transformers. Firstly, the original top-oil temperature series was preprocessed. Secondly, a parallel forecasting model was implemented by LSTM and SA, and fused the parallel forecasting features using GRU. Finally, the forecasting results were obtained after adjustment by the fully connected layer. The experimental results show that the proposed method is superior to other existing single forecasting models in UHV converters top-oil temperature forecasting.