Abstract: In order to address the problem of long time-consuming calculation of the steady temperature rise of oil-immersed transformer winding, this paper proposes a fast calculation method based on POD-DNN reduced-order model (ROM). First, the snapshot matrix is constructed through the full-order model (FOM) of the steady temperature rise of transformer winding, and the modes and modal coefficients of the physical system are obtained based on proper orthogonal decomposition (POD). Next, a deep neural network (DNN) agent model between the working conditions and modal coefficients is established to solve the limitations of low solving efficiency of nonlinear terms and strong dependence of control equations in POD, and design network regularization strategies to avoid overfitting of the model under small samples. Finally, the modal coefficients predicted by DNN agent model are combined with the corresponding POD modes to reconstruct the temperature field of winding. It is verified that the winding temperature rise results solved by POD-DNN are very close to the Fluent simulation and experimental measurements, and the computational efficiency is improved by 247478 and 23056 times compared with the FOM and Fluent simulation, respectively. This algorithm can provide technical support for online monitoring, operation maintenance, and insulation design of transformers.