Extended Physics-Informed Neural Networks for Parameter Identification of Switched Mode Power Converters With Undetermined Topological Durations

It is challenging to perform high-precision parameter identification for switched mode power converters with undetermined topology duration because the physical information, such as switching instants and circuit state variables at topology transitions is critical for realizing this goal. In conventional physics-based solutions, additional measurement circuits are required to compensate for the absence of the unknown physical information at topology transitions, otherwise accuracy has to be compromised. To avoid the undesired additional hardware, an extended PINN (e-PINN), which integrates a pseudolabel generation network into the piecewise PINN, is proposed. This network can precisely identify key system parameters, along with the duration of each topology and system states at topology transitions. The effectiveness of the e-PINN is experimentally validated on a buck converter operating in discontinuous conduction mode (DCM), which, is a basic case of the power converter having undetermined topology duration. Compared with the traditional physics-based parameter identification methods for DCM buck converter, e-PINN can precisely estimate system parameters without necessitating high-frequency sampling or zero-current detection circuits, which increase the cost, volume, and safety risk. Besides, it can operate without disrupting system operation.