Probabilistic power flow (PPF) plays a critical role in the analysis of power systems. However, its high computational burden makes practical implementations challenging. This paper proposes a model-based deep learning approach to overcome these computational challenges. A deep neural network (DNN) is used to approximate the power flow calculation process, and is trained according to the physical power flow equations to improve its learning ability. The training process consists of several steps: 1) the branch flows are added into the objective function of the DNN as a penalty term, which improves the generalization ability of the DNN; 2) the gradients used in back propagation are simplified according to the physical characteristics of the transmission grid, which accelerates the training speed while maintaining effective guidance of the physical model; and 3) an improved initialization method for the DNN parameters is proposed to improve the convergence speed. The simulation results demonstrate the accuracy and efficiency of the proposed method in standard IEEE and utility benchmark systems.