Unified dual-PINN solution for DC-DC power converter modeling and control with fast piecewise CPL sensing

• A novel three-echelon iPINN-DPINN-MPC control framework is proposed for DC-DC buck converters. • The inverse and direct PINN networks are trained offline using converter physics and representative data. • Real-time co-simulation with Arduino Due and MATLAB/Simulink validates hardware feasibility. • The proposed controller achieves superior voltage tracking and adaptive response to load changes. • Experimental results show less than 1% deviation compared to simulation, confirming robustness. Constant-power loads (CPLs) impose strong nonlinearities on buck converters operating in continuous conduction mode (CCM), making stability and control highly sensitive to load power and passive parameters. This work proposes a three-stage inverse-to-forward physics-informed framework for online identification and predictive control. In the first stage, an inverse PINN (iPINN) jointly estimates the piecewise-constant CPL power and the passive parameters ( L, C ) from voltage–current trajectories by enforcing averaged CCM dynamics and regularization terms, enabling reliable online identification during rapid transients. In the second stage, a direct PINN (DPINN) is trained using the identified parameters to construct a stable grey-box surrogate that embeds the buck conservation laws and generalizes across operating points. In the third stage, this physics-aware surrogate is integrated into a model predictive controller (MPC) to perform short-horizon duty-ratio optimization with accurate state forecasts. A key contribution of this work is a dual-PINN architecture—combining inverse estimation with forward physics-consistent prediction—that forms a unified identification–prediction–control pipeline. This integrated structure significantly reduces model–plant mismatch and enhances robustness to CPL steps and parameter drifts compared with conventional model-based or purely data-driven MPC schemes. Simulation results demonstrate (i) low-error parallel estimation of CPL and ( L, C ), (ii) stable convergence of the forward surrogate, and (iii) improved closed-loop MPC performance, establishing a practical pathway from iPINN-based identification to real-time control.