Abstract
Off-grid green hydrogen production couples photovoltaic generation, battery storage, and proton exchange membrane (PEM) electrolysis on a shared DC bus. Advanced nonlinear controllers are routinely proposed for this coupling, yet rarely compared against simpler baselines under a framework covering aging and economics, and rarely decomposed to show which element of the design produces the benefit. We compare a PI cascade, input–output feedback linearization (IO-FL), and interconnection and damping assignment passivity-based control (IDA-PBC) on a seven-state averaged model of a photovoltaic, battery, and PEM electrolyzer system sized for northern Chile, together with three ablation variants: a PI cascade with a model-based power-balance feed-forward, IO-FL with integral action, and IDA-PBC retuned to the common damping ratio. All are tuned to the same closed-loop bandwidth and tested under cloud drop-outs, 200 paired Monte Carlo runs with ±20–30% plant perturbations, a second Monte Carlo set in which the controllers operate with nominal models, measurement bias and sensor errors, a representative day with supervisory start/stop operation, published degradation models, and a levelized cost of hydrogen (LCOH) analysis. On the slow bus-voltage manifold, the IDA-PBC law coincides with IO-FL augmented by an integrator, and the regulation benefit of both is explained by the power-balance feed-forward: adding that single term to the PI cascade lowers the peak DC bus deviation from 4.57% to 0.65%, below the 0.88% of IDA-PBC, and it does so in every Monte Carlo realization. When the controllers operate with nominal models and corrupted measurements, the four variants that combine the feed-forward with integral action become indistinguishable and keep the ±2% target in every realization, while the canonical IO-FL misses it in three realizations out of four. The three design routes, however, are indistinguishable on cost and component life: under supervisory operation, the battery throughput and the PEM start/stop count are set by the energy management layer, and the six controllers give the same LCOH to the fourth decimal place. For the architecture, supervisory strategy, and operating scenarios investigated here, the case for a particular control law rests on regulation quality, robustness to model error, and the availability of a stability certificate, not on longer component life or cheaper hydrogen.