Abstract: In the research field of twin modeling and simulation of oil-immersed transformers, there is a significant technological bottleneck in the insufficient capability for multi-physics coupling analysis, particularly in the efficient computation of the temperature distribution within transformer windings. This paper proposes a rapid XGBoost-based computation method that incorporates spatial feature relationships, aiming to facilitate the fast calculation of the temperature field in transformer windings. Firstly, a temperature field simulation is conducted for a 35 kV oil-immersed self-cooled transformer within the framework of multi-physics coupling theory. Through accurate numerical simulations and precise definition of physical parameters, a winding temperature dataset is constructed that accurately reflects the internal heat transfer processes within the transformer. Subsequently, the dataset is used as input for model training, and the results are compared with those obtained from simulation experiments and other model training approaches. Finally, temperature rise experiments are performed to verify the validity of both the simulation results and the model's predictions. During the experiment, the method proposed in this paper reduces the multi-physics simulation time from hours to mere seconds. Under identical operating conditions, the average prediction error for the high-voltage winding is 0.22 ℃, while the average error for the low-voltage winding is 1.72 ℃. The results indicate that the XGBoost-based rapid computation method, which considers spatial feature relationships, effectively enables fast calculation of transformer winding temperatures, thereby meeting the real-time and accuracy requirements for digital twin modeling and simulation.