As a critical unconventional oil and gas resource within the global energy framework, heavy oil has garnered significant attention for its development efficiency. Although steam flooding technology has improved the efficiency of heavy oil production, the phenomenon of steam breakthrough negatively impacts thermal efficiency and reservoir development. Traditional prediction methods have shown inadequate precision and delayed response when dealing with long-term oilfield time series data. Data from 13 steam flooding well groups in the Shengli oilfield heavy oil block were utilized. An innovative approach was adopted, using the instantaneous temperature ratio between production and injection wells as an indicator of steam breakthrough time. Pearson correlation coefficient analysis was employed to select key factors related to steam breakthrough time. Based on these factors, a deep learning model built on the Transformer architecture was developed, achieving accurate predictions of the instantaneous temperature ratio. The predictions closely aligned with oilfield observation data, demonstrating higher prediction accuracy and stability compared to traditional long short-term memory(LSTM) models. The research results not only provide a new perspective for the precise prediction of steam breakthrough time in heavy oil reservoirs but also further validate the extensive potential of deep learning technology in oilfield development applications, supporting the construction of intelligent oilfield management and decision support systems.