In order to solve the deficiencies of traditional finite element methods in terms of computational efficiency and real-time interactivity in digital twin systems, this paper proposes a deep neural network-based agent modeling fast computation method for rapid computation of transformer bushing flange stress field. The study first establishes a finite element model of a 110 kV transformer bushing, and then combines the electro-thermal-force coupling effect with end loads and conductor currents as inputs, and utilizes a deep neural network architecture to train the stress field data. Finally, the finite element simulation results were compared with the proxy model training results, which showed an average accuracy of 95.8% and significantly reduced the computation time from 330 days to less than 4 days. The study points out accuracy problems in extreme parameter ranges, which may be caused by insufficient data or improper choice of activation function. The method provides a reference for rapid multi-physics field simulation and digital twin reconstruction.