Neural Electromagnetic Transients Program

This paper devises a neural electromagnetic transients program (NeuEMTP), an unsupervised, physics-informed learning approach to numerical-integration-free EMTP solutions. The main contributions lie in: (1) a learning-based NeuEMTP architecture to simultaneously generate the electromagnetic states at all desired time steps, making the step-by-step integration unnecessary; (2) an unsupervised, physics-informed training procedure to realize the NeuEMTP functionality without requiring any EMTP trajectories beforehand; (3) an EMTP-oriented-neural-network (EMTPNet) accompanied with a novel activation function Act_mix to enable efficient extrapolations of diverse oscillation modes under arbitrary frequencies. Case studies sys-tematically verify that NeuEMTP generates high-fidelity EMTP trajectories without involving any numerical integration before or during the training process, and is promising to achieve faster-than-real-time EMTP simulations on the off-the-shelf computers.