Answers are grounded in this paper and cite their sources.
Co-locating hydrogen production and aquaculture on floating offshore wind platforms requires dispatch policies that jointly optimize energy trading, electrolyzer longevity, and cage water quality—objectives that scalarized reinforcement learning struggles to reconcile. We develop a conservation-closed reduced-order model of a 20 MW-class floating wind system, driven by ERA5 North Sea reanalysis and cross-checked against OpenFAST, and cast dispatch as a Markov decision process. Scalarized twin delayed deep deterministic policy gradient (TD3) agents achieved nominally higher revenue than rule-based control (+0.033 M€; Hedges’ g = 0.94) but converged to a degenerate, hydrogen-abandoning equilibrium in 7% of independent runs (2/27). A hierarchical TD3 agent (H-TD3) with an annealed logarithmic Nash-bargaining reward showed no collapse across 20 seeds (Fisher’s exact test, p = 0.32), reached 95% of its final physical-objective level at the earliest observable episode in every seed, and exhibited an order-of-magnitude smaller plateau dispersion than scalarized baselines, with revenue statistically indistinguishable from rule-based control (1.216 ± 0.040 M€). Over 1000 h, H-TD3 increased cage biomass by 22.6% and yielded an electrolyzer-subsystem hydrogen cost of €3.05/kg (excluding electricity, wind-farm, and aquaculture costs), or €6.45/kg at €60/MWh wind electricity. These results frame multi-objective offshore dispatch as a risk-control problem: Nash-bargaining rewards produced fewer collapses than linear scalarization, though the difference was not significant.
No takes yet. Share an insight, caveat, or question.
Lyu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: