Abstract: Active Magnetic Bearing (AMB) is a core part to achieve non-contact support in flywheel energy storage system. Its electromagnetic force to the rotor is related to several nonlinear factors, but a conventional linear model based on "two step linearization" removes nonlinear information. It is not able to express the broadband spectrum working condition, and has negative influence to design accuracy of control subsystem. To solve this problem, a new nonlinear model base on alternative mapping concept was developed. An AMB finite element model was built. Its parametric calculation data were used to train a BP neural network. Weight and threshold matrix from this training was used to build a zero-dimensional nonlinear dynamic model, which can describe relationship from rotor displacement and winding current to forces to rotor. This model was implemented in Matlab/Simulink. Calculation results show there are complex nonlinear characteristics in an AMB, including four order terms and cross terms. A comparison with linear model results shows that this new model is more accurate in AMB's nonlinear characters. This model is an important support to system simulation and control subsystem development in flywheel energy storage system.