Topology-Cognizant Machine Learning for Multiparameter Estimation and Digital Twin of LLC Resonant Converters

System parameters, particularly those within the resonant tank, fundamentally govern the converter’s dynamic behavior and operational performance of LLC resonant converters. Accurate estimation of these system parameters is essential for enabling real-time condition monitoring, high-efficiency operation, and advanced model-based control of LLC resonant converters. However, due to strong parameter coupling, high system nonlinearity, and unobservable topology transitions, performing accurate multi-parameter estimation for LLC converters remains considerably challenging, and has rarely been effectively solved in existing works. To address this issue, this study proposes a topology-cognizant machine learning (TCML), which explicitly embeds the converter’s discrete-time topology-dependent dynamics into a neural network architecture. In contrast to conventional PINN, TCML introduces a semi-supervised data reconstruction mechanism to infer system states at unobservable topological transitions, enabling consistent supervision across discontinuous dynamic regimes. This physics-guided structure compels the network to jointly satisfy observed data consistency and converter-specific physical laws, thereby achieving high-accuracy multi-parameter identification under a data sampling configuration compatible with commonly used microcontroller units (MCUs). Furthermore, the proposed method offers additional advantages, including reduced reliance on accurate parameter pre-estimation, nanosecond-scale digital twin capability, and online implementation, The effectiveness is verified through both simulations and experiments.