Self-Supervised Learning For Frequency-Constrained Scheduling of Off-Grid Renewable Power-To-Hydrogen Systems

Renewable power-to-hydrogen (ReP2H) systems have emerged as an effective solution for local renewable consumption, but their economic scheduling remains challenging. On one hand, the nonlinear efficiency of electrolyzers introduce strong nonconvexities into the optimization problem. On the other hand, ensuring reliable operation of low-inertia ReP2H systems requires enforcing nonlinear transient frequency-security constraints, which significantly increase computational complexity. To address these challenges, this paper proposes a novel frequency-constrained scheduling framework tailored for ReP2H systems based on selfsupervised learning (SSL). A deep neural network (DNN)-based frequency dynamics surrogate (FDS) is first developed and trained through an accelerated structure function-enhanced training strategy, which balance the high predictive accuracy and training efficiency under limited budgets of labeled data generation. The trained FDS is then embedded within an SSL-based primal–dual learning SSL (PDL-SSL) optimization framework, where the primal and dual DNNs are jointly trained to learn a scheduling proxy that preserves the nonlinear dynamics of electrolyzers and allows end-to-end training without relying on pre-solved optimal decisions. Finally, case studies on 14-node and 33-node ReP2H systems demonstrate that the proposed method achieves nearoptimal scheduling results while satisfying frequency-security constraints with substantially reduced computation time.