Multi-objective optimization framework for reconfigurable PV–PEM electrolyzer using machine learning surrogates

Optimization and electrical control of photovoltaic (PV) and proton exchange membrane (PEM) electrolyzer systems are critical for efficient, scalable hydrogen production. Static coupling methods employ fixed electrical interconnections between PV arrays and the PEM stack, which cannot adapt to fluctuating irradiance or temperature. To overcome this, we propose an adaptive reconfiguration optimization strategy that dynamically adjusts PV electrical configurations and co-optimizes PEM operating temperatures. This allows the system to emulate maximum power point tracking (MPPT) and optimize power management without additional power electronics. Using NSGA-II and CRITIC-based objective weights, we identified feasible solutions for multi-objective optimization. Building upon this, a smart hybrid energy system approach combining machine learning with real-time optimization is introduced to enhance adaptability. Specifically, a Random Forest regression model is trained to predict optimal configurations and performance metrics based on environmental inputs. This approach significantly reduces reliance on computationally intensive iterative optimization, enabling near-instantaneous control while introducing only controlled approximations. Results show that the hybrid approach decreases computation time by 6-fold while maintaining system performance across three diverse locations. These findings highlight the potential for accurate sizing and reliable control of directly coupled PV/PEM systems with substantially reduced computational cost.