To address the issue of system-level electromagnetic compatibility, a new method of predicting electromagnetic interference of complex systems based on artificial neural network (ANN) reverse model is proposed in this paper. Firstly, the electric field radiated emission (RE) of single equipment is measured. The training data of system-level RE are obtained by simulation based on the equivalence principle of radiated emission. Frequency, RE and coordinate of each single equipment are selected as the input variables, and the system-level RE is the output variable. A reverse model of the three-layer back-propagation (BP) ANN with Levenberg-Marquardt (LM) algorithm is established by exchanging the input–output variables. The alternative ANN with minimum validation error is searched as the ultimate ANN. The numerical root-finding algorithm (regular-falsi method and conjugate gradient method) are adopted to calculate the RE of multi equipments. The results show that the validation error of this reverse model is significantly improved compared to the traditional three-layer LM-BP ANN. Especially, the ANN reverse model based on conjugate gradient method reduces the validation error from 0.4159% to 0.0997%. This method is independent of complex ANN structures, and improves simulation accuracy with limited training data, which provides a new efficient and feasible solution for electromagnetic compatibility evaluation of electronic information platforms such as ships, satellites, and aircrafts.