An analytically separated neuro-space mapping (Neuro-SM) model of power transistors is proposed in this paper. Two separated mapping networks are introduced into the new model to improve the characteristics of the DC and AC, avoiding interference of the internal parameters in neural networks. Novel analytical formulations are derived to develop effective combinations between the mapping networks and the coarse model. In addition, an advanced training approach with simple sensitivity analysis expressions is proposed to accelerate the optimization process. The flexible transformation of terminal signals in the proposed model allows existing models to exceed their current capabilities, addressing accuracy limitations. The modeling experiment for the measurement data of laterally diffused metal-oxide-semiconductor transistors demonstrates that the novel method accurately represents the characteristics of the DC and AC of transistors with a simple structure and efficient training process.