Physics-Seeded Cascaded Refinement-Inference Network for Timeseries Prediction of Dual-Valve Heave Compensators with Parameter Variations

The Dual-valve Heave Compensator (DHC) is a semi-active offshore system for motion decoupling, yet high-performance control is constrained by the computational cost of hysteretic coupled models under simultaneous variations in external, internal, and control parameters. To address such bottleneck, in this work a Physics-Seeded Cascaded Refinement-Inference Network (PSCRIN) is proposed. PSCRIN is a three-level architecture: (i) a Numerical-Solver Physics-Seed sub-Network (NSPSN) that rapidly produces coarse displacement outputs via a shallow, physics-seeded structure; (ii) a Refinement Deep Feedforward sub-Network (RDFFN) that refines these outputs with configurable memory dependence; and (iii) a State Inference sub-Network (SIN) that efficiently infers dynamic system states using a shallow design. High fidelity simulations demonstrate that PSCRIN accurately reproduces hysteresis-inclusive DHC displacement and state responses, achieving consistently high statistical performance and providing a practical foundation for broader control applications in offshore operations.