The neural Lyapunov method is an effective tool for assessing the transient stability region of power systems. However, existing neural Lyapunov methods for constructing Lyapunov functions for power systems have issues such as long training times, overfitting, and strong conservatism in stability region estimation. This paper proposes a Lyapunov function training structure based on residual neural networks (Resnet), introducing bypass branches in the feedforward neural network to alleviate the overfitting problem and improve training speed. The method has been validated on systems with virtual synchronous generator-controlled converters and phase-locked loop-controlled converters connected to the grid. Experimental results show that compared to other neural Lyapunov methods, this method can construct effective Lyapunov functions more quickly and estimates the system’s stability region with less conservatism.