Learning-Assisted Optimization of Planning and Operation in Wind-Hydrogen Power Systems

Abstract

This paper proposes an integrated planning-operation framework coupling data-driven wind modeling with bi-layer optimization for hydrogen-integrated power systems. To address uncertainty, a stacked Long Short-Term Memory (LSTM) network is used to generate year-long hourly offshore wind profiles, preserving multiscale temporal dynamics. These profiles are embedded into a bi-layer optimization structure: an upper-layer Particle Swarm Optimization (PSO) determines the siting and sizing of hydrogen-to-power assets, while a lower-layer Mixed-Integer Linear Programming (MILP) model ensures operational feasibility under network, storage, and capacity constraints. The framework is validated on a modified IEEE RTS-24 system with large-scale offshore wind. Results demonstrate that coordinated fuel cell and hydrogen-fired gas turbine deployment significantly improves performance, reducing wind curtailment by 48.5% and load shedding by 96.6%, while cutting annual costs by $ 20.06 M. Despite hydrogen's modest generation share, it serves as a critical flexibility and reliability enabler. Ultimately, this study demonstrates how neural-network-based uncertainty modeling effectively informs strategic, system-level investment decisions in modern power grids.