Large-scale electrical equipment typically involves millions of nodes

leading to prohibitive computational costs in three-dimensional (3-D) nonlinear transient magnetic field simulations and making real-time simulation of electromagnetic performance difficult to achieve. This paper proposes a fast calculation method for nonlinear transient magnetic fields based on cluster-enhanced proper orthogonal decomposition (CPOD) and stepwise multi-branch bidirectional long short-term memory (SMB-BiLSTM) reduced-order modeling. A reduced snapshot matrix is constructed via a multi-stage K-means spatial node reduction strategy to achieve efficient feature extraction of the magnetic fields in large-scale electrical equipment

and the dominant modes and their corresponding coefficient vectors are obtained. In accordance with the 3-D transient characteristics

an SMB-BiLSTM network is established to calculate the temporal coefficient vectors under different excitation waveforms

thus the complex field solving problem is transformed into a prediction of the optimal linear combination of different modes

ultimately reconstructing the precise magnetic field distribution. The efficiency of the proposed reduced-order model is verified by establishing fast calculation models for the TEAM Problem 24 (a benchmark for nonlinear transient fields in rotating machines) and a 1 000 kVA amorphous alloy distribution transformer

and the results and computation time are compared with those of the full-order models. The results demonstrate that the proposed method can be adopted to maintain high calculation accuracy (with average relative errors of 0.79% and 1.72%

respectively) and significantly improve the calculation efficiency of 3D nonlinear transient magnetic fields

achieving a speedup ratio of over 10 000 times.